0% found this document useful (0 votes)

131 views

Analytics Optimization With Columnstore Indexes in Microsoft SQL Server

Uploaded by

yolode

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

131 views

Analytics Optimization With Columnstore Indexes in Microsoft SQL Server

Uploaded by

yolode

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 285

Analytics

Optimization with
Columnstore
Indexes in Microsoft
SQL Server
Optimizing OLAP Workloads
—
Edward Pollack
Analytics Optimization
with Columnstore Indexes
in Microsoft SQL Server
Optimizing OLAP Workloads

Edward Pollack
Analytics Optimization with Columnstore Indexes in Microsoft SQL Server:
Optimizing OLAP Workloads
Edward Pollack
Albany, NY, USA

ISBN-13 (pbk): 978-1-4842-8047-8 ISBN-13 (electronic): 978-1-4842-8048-5

https://doi.org/10.1007/978-1-4842-8048-5
Copyright © 2022 by Edward Pollack
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
Trademarked names, logos, and images may appear in this book. Rather than use a trademark symbol with
every occurrence of a trademarked name, logo, or image we use the names, logos, and images only in an
editorial fashion and to the benefit of the trademark owner, with no intention of infringement of the
trademark.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not
identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to
proprietary rights.
While the advice and information in this book are believed to be true and accurate at the date of publication,
neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or
omissions that may be made. The publisher makes no warranty, express or implied, with respect to the
material contained herein.
Managing Director, Apress Media LLC: Welmoed Spahr
Acquisitions Editor: Jonathan Gennick
Development Editor: Laura Berendson
Coordinating Editor: Jill Balzano
Cover image designed by Freepik (www.freepik.com)
Distributed to the book trade worldwide by Springer Science+Business Media LLC, 1 New York Plaza,
Suite 4600, New York, NY 10004. Phone 1-800-SPRINGER, fax (201) 348-4505, e-mail orders-ny@springer-
sbm.com, or visit www.springeronline.com. Apress Media, LLC is a California LLC and the sole member
(owner) is Springer Science + Business Media Finance Inc (SSBM Finance Inc). SSBM Finance Inc is a
Delaware corporation.
For information on translations, please e-mail [email protected]; for reprint,
paperback, or audio rights, please e-mail [email protected].
Apress titles may be purchased in bulk for academic, corporate, or promotional use. eBook versions and
licenses are also available for most titles. For more information, reference our Print and eBook Bulk Sales
web page at http://www.apress.com/bulk-sales.
Any source code or other supplementary material referenced by the author in this book is available to
readers on GitHub at https://github.com/Apress/analytics-optimization-w-columnstore-indexes-
in-microsoft-sql-server.
Printed on acid-free paper
For Theresa, Nolan, and Oliver, without whom none of this
would be possible.
Table of Contents
About the Author�� xi

About the Technical Reviewer�� xiii

Acknowledgments��xv

Introduction��xvii

Chapter 1: Introduction to Analytic Data in a Transactional Database�� 1

Where Should Analytic Data Reside?�� 1
Analytic Data Size�� 2
Analytic Data Structure�� 4
Analytic Data Sources�� 5
Data Warehouse�� 6
Unstructured Data�� 6
Third-Party Analytics Software�� 6
OLAP Data in an OLTP Table in a Transactional Database�� 7
OLAP Data in an Analytic Table in a Transactional Database�� 8

Chapter 2: Transactional vs. Analytic Workloads�� 11

Transactional Data�� 11
Analytic Data�� 14
The Need for Two Systems�� 18
Building Better Analytic Data Structures�� 19

Chapter 3: What Are Columnstore Indexes?�� 21

The Limits of Transactional Data Storage�� 21
Introducing Columnstore Indexes�� 24

v
Table of Contents

Benefits of Columnstore Indexes in SQL Server�� 26

Native Analytic Data in SQL Server�� 26
Scalability�� 27
Exceptional Compression�� 29
Faster Analytic Reads�� 30
Faster Data Loads�� 30

Chapter 4: Columnstore Index Architecture�� 33

Sample Data�� 33
Rowgroups and Segments�� 36
The Delta Store�� 40
The Delete Bitmap�� 41
Nonclustered Columnstore Index Architecture�� 42
Physical Data on Pages�� 43
Summarizing Differences�� 47

Chapter 5: Columnstore Compression�� 49

Basics of Columnstore Compression�� 49
Columnstore Compression Algorithms�� 51
Value Encoding�� 51
Dictionary Encoding�� 54
Should String Data Be Normalized?�� 60
Row Order (Vertipaq) Optimization�� 62
Other Compression Algorithms�� 65
Columnstore Archive Compression�� 68
The Compression Life Cycle�� 71

Chapter 6: Columnstore Metadata�� 73

Available Columnstore Metadata�� 73
Rowgroup Metadata�� 73
Segment Metadata�� 75
Rowgroup Physical Metadata�� 79

vi
Table of Contents

Rowgroup Operational Statistics�� 84

Columnstore Index Memory Usage�� 89
Internal Columnstore Index Objects�� 93

Chapter 7: Batch Execution�� 97

Row Mode Execution�� 97
Batch Mode Execution�� 100
How Does Batch Mode Work?�� 105
Batch Mode vs. Row Mode Performance�� 107

Chapter 8: Bulk Loading Data�� 111

Bulk Load Processes Explained�� 111
Bulk Loading into Columnstore Indexes�� 112
Performance of Bulk Loading into Columnstore Indexes�� 114
Trickle Insert vs. Staged Insert�� 118
Other Data Load Considerations�� 119
Drop Nonclustered Indexes During Data Loads�� 119
Columnstore Reorganize Operations with Each Data Load�� 120
Summary�� 122

Chapter 9: Delete and Update Operations�� 123

The Cost of Modifying Data�� 123
Delete Operations�� 125
Update Operations�� 128

Chapter 10: Segment and Rowgroup Elimination�� 137

Segment Elimination�� 137
Rowgroup Elimination�� 142
Combining Segment and Rowgroup Elimination�� 154

Chapter 11: Partitioning�� 157

Maintain Hot/Warm/Cold Data�� 157
Faster Data Movement/Migration�� 158

vii
Table of Contents

Partition Elimination�� 159

Database Maintenance�� 160
Partitioning in Action�� 161
Partitioning Guidelines�� 175
Partition and Rowgroup Sizing�� 175
Partition Column Choice�� 176
Storage Choice�� 176
Additional Benefits�� 177

Chapter 12: Nonclustered Columnstore Indexes on Rowstore Tables�� 179

Use Rowstore Indexes�� 180
Separate OLAP and OLTP Processes�� 183
Nonclustered Columnstore Indexes�� 185
Managing Hot, Warm, and Cold Transactional Data�� 187
Compression Delay�� 188
Filtered Nonclustered Columnstore Indexes�� 195
Code Changes�� 199
Vertipaq Optimization for Nonclustered Columnstore Indexes�� 199
Testing Nonclustered Columnstore Indexes�� 203
Don’t Forget to Drop Unneeded Indexes!�� 204

Chapter 13: Nonclustered Rowstore Indexes on Columnstore Tables�� 207

Using Nonclustered Rowstore Indexes�� 207
Enforcing Constraints�� 212
Filtered Nonclustered Rowstore Indexes�� 214
Enabling Vertipaq Optimization�� 216
Add Filters to Nonclustered Rowstore Indexes�� 216
Perform Periodic Index Maintenance�� 216
Indexed Views�� 218
Compression for Nonclustered Rowstore Indexes�� 220
Nonclustered Rowstore Index Guidance�� 222

viii
Table of Contents

Chapter 14: Columnstore Index Maintenance�� 225

What Causes Fragmentation?�� 225
How Much Fragmentation Is Too Much?�� 228
Quantifying Deleted Rows�� 228
Detailing Unordered Data�� 230
The No-Maintenance Columnstore Index�� 236
Columnstore Reorganize�� 237
Reorganize to Remove Delta Rowgroups�� 239
Columnstore Rebuild�� 242
Columnstore Reorder and Rebuild�� 245
Columnstore Index Maintenance by Partition�� 246
Index Maintenance in Nonclustered Columnstore Indexes�� 247

Chapter 15: Columnstore Index Performance�� 249

Columnstore Metadata Reads�� 249
Columnstore Data Reads�� 253
Memory Sizing�� 254
Dictionary Size and Dictionary Pressure�� 257
Normalizing Wide Columns�� 258
Add or Change the Columnstore Sorting Column�� 259
Partitioning�� 259
Columnstore Indexes on Temporary Tables�� 260
Memory-Optimized Columnstore Indexes�� 264
Demonstrating Memory-Optimized Columnstore Indexes�� 264
Optimization Strategies�� 274

Index�� 275

ix
About the Author
Edward Pollack has over 20 years of experience in database and systems administration,
architecture, and development, becoming an advocate for designing efficient data
structures that can withstand the test of time. He has spoken at many events, such as
SQL Saturdays, PASS Data Community Summit, Dativerse, and at many user groups and
is the organizer of SQL Saturday Albany. Edward has authored many articles, as well as
the book Dynamic SQL: Applications, Performance, and Security, and a chapter in Expert
T-SQL Window Functions in SQL Server. His first patent was issued in 2021, focused on
the compression of geographical data for use by analytic systems.
In his free time, Ed enjoys video games, sci-fi and fantasy, traveling, and baking.
He lives in the sometimes-frozen icescape of Albany, NY, with his wife Theresa and sons
Nolan and Oliver, and a mountain of (his) video game plushies that help break the fall
when tripping on (their) toys.

xi
About the Technical Reviewer
Borbala Toth-Apathy is a database professional with nearly 20 years of experience in
the IT field. She has an MSc in Computer Science – unsurprisingly, as she’s been writing
code from a young age.
Specializing in databases after her degree, nowadays, she is designing and building
data warehouses. Her secret passion is data analytics.
In her free time, she likes calisthenics or yoga, nature, puzzles, and crafts.

xiii
Acknowledgments
A big shout-out to the SQL Server community and the many speakers, organizers,
colleagues, and others who have supported and advised me over the years and provided
opportunities to grow both personally and professionally.
A book doesn’t happen without dedicated reviewers and editors. Speakers do not
get the opportunity to share their knowledge and grow without organizers, sponsors,
and volunteers that provide those chances. Articles are not written without companies,
editors, and services to support and encourage that work.
To everyone that has provided those opportunities to be a part of your events,
publications, and groups: Thank you!!!

xv
Introduction
Analytic data is an ever-present challenge for developers, analysts, and data
professionals. Its rapid growth coupled with the constant organizational need for speedy
answers to complex questions leads to the question of how massive amounts of data can
be stored efficiently enough to allow for real-time analytics.

What Is This Book?

This book uses columnstore indexes as a tool to address the challenge of storing different
types of analytic data efficiently. Applications such as data warehousing, logging,
reporting, analytics, and data science can all be solved by effective use of columnstore
indexes. In addition, real-time operational analytics can be accomplished with this tool.
Analytics Optimization with Columnstore Indexes in Microsoft SQL Server is
structured into three functionally different sections:

• Chapters 1 and 2: Introduction to analytic data and its challenges

• Chapters 3 and 4: Columnstore index basics and architecture

• Chapters 5–15: Deep dives into features, architecture, and best

practices

The goal is to provide enough theory, architecture, use cases, and code to enable the
immediate use of this feature in solving real-world problems.
Each chapter tackles a specific feature or data architecture need from a high-level
overview, then delving into extensive detail. While not all professionals may need to
know precisely how columnstore indexes compress their data, having that knowledge
available will provide value in the future when the need to improve performance arises.
Most examples in this book are derived from data in the WideWorldImportersDW
and WideWorldImporters databases. On their own, they do not provide enough data to
allow for meaningful examples of analytic data. To augment this, new tables are created,
copied, and modified to allow for objects that are large and interesting enough to be
useful in demonstrating columnstore indexes. While real-world tables can easily contain

xvii
Introduction

tens of billions of rows, the examples in this book settle for about 25 million rows. This
provides a healthy compromise between effective test data and tables that are so large
that building them takes a very long time.

Intended Audience
This book is intended for data professionals, architects, and developers as a tool to
tame reporting and analytic data quickly, inexpensively, and effectively. Topics in this
book range from introductory to quite advanced, allowing it to be used as a tool by both
beginners and those with years of data architecture experience under their belts.
Those who have little experience with columnstore storage technologies will learn
everything needed to implement this feature for the first time and begin realizing
its benefits. Data experts can use the feature and architecture details, as well as
demonstrations provided throughout this book as a tool to optimize existing data
structures.
Columnstore indexes are not only for reporting and analytic data. They can also be
used to table the complex organizational need for real-time operational analytics. When
analytic and transactional workloads mix, the risk exists for latency, blocking, locking,
and deadlocking. Developers and data professionals that work with mixed workloads
can benefit greatly from the use of columnstore indexes to help cover each workload
effectively. In addition to improving performance, these efforts can avoid the need to
purchase or architect new systems to manage portions of a mixed workload.

Why Columnstore Indexes?

There are many solutions available that help tame analytic data, but most require
investing heavily in new hardware, software, technologies, and architecture. Not all
organizations can (or should) feel forced into investing immense resources in new
solutions unless there is a truly compelling reason to do so.
Columnstore indexes can be implemented in SQL Server using existing hardware
and licenses, allowing for analytic data solutions to be realized without a significant
investment of time and resources. This allows for easier QA and testing of the feature
without the need to commit to it up front.

xviii
Introduction

In addition to providing a robust solution, columnstore indexes are based on mature

technology that has underpinned many other data storage products (such as PowerPivot
and SSAS) for years. Despite their maturity, columnstore indexes in SQL Server evolve
and improve with each new version that is released.

Contacting the Author

This book is incomplete without readers, and this experience is only truly meaningful
when I am available to answer questions and accept feedback. The thoughts that are
shared with me at conferences, industry events, and through electronic means provide
ways to grow and improve, as well as ideas worthy of consideration.
Feel free to contact me about this book, its topic, or related thoughts via Twitter
(@EdwardPollack) or email ([email protected]).

xix
CHAPTER 1

Introduction to Analytic
Data in a Transactional
Database
Analytic data, by its nature, can be large and challenging to maintain and can grow
quickly over time. Similarly, its use increases with time as analysts and data scientists
find more ways to crunch it. There is a great convenience to having analytic data in close
proximity to its underlying transactional sources. Utility is also gained by choosing a
location for analytic data that can withstand the test of time, thus avoiding the need for
costly migrations if the data is unable to scale appropriately.

Where Should Analytic Data Reside?

Until the advent of columnstore indexes, storing analytic data in a transactional database
in SQL Server was cumbersome and required tricks and compromises to make it
performant enough for the rigors of analytics. Other technologies and services were
incorporated as a way to work around these limitations.
There are a variety of places that analytic data can be stored, such as

• Data warehouse (SQL Server Analysis Services, RedShift, etc.)

• Unstructured data (data lake using Hadoop, Hive, etc.)

• Third-party analytics software

• OLTP tables in a relational/transactional database

• OLAP tables in a relational/transactional database

1
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_1
Chapter 1 Introduction to Analytic Data in a Transactional Database

Each of these options has advantages and disadvantages, but before diving into
specifics, it is important to review analytic data and how it looks.
Note that OLAP stands for “online analytic processing” and references analytic
tables and data, whereas OLTP stands for “online transactional processing” and refers to
transactional tables and data.

Analytic Data Size

Typical analytic data has significantly higher row counts than its source transactional
data. This may be due to tracking many metrics in extensive detail, or it may be due to
tracking those details in repeated samples over a long period of time. If a transactional
log table adds a million rows per day, then 365 million rows will be captured per year.
Figure 1-1 shows the dramatic effect that time can have on data and how a small data set
can grow very quickly when sampled regularly.

Figure 1-1. Illustration of the growth of data when the time dimension is added

Analytic data can also be quite wide. Even a small table can have many calculated
metrics derived from it, allowing its OLAP counterpart to contain significantly more
columns. For example, consider the hypothetical sales order table in Listing 1-1.

2
Chapter 1 Introduction to Analytic Data in a Transactional Database

Listing 1-1. An Example of a Transactional Sales Order Table

CREATE TABLE dbo.SalesOrder

(      SalesOrderId INT NOT NULL IDENTITY(1,1) CONSTRAINT PK_SalesOrder
PRIMARY KEY CLUSTERED, ProductDetailList INT NOT NULL CONSTRAINT
FK_SalesOrder_ProductDetail FOREIGN KEY REFERENCES dbo.ProductDetail,
       CustomerId INT NOT NULL CONSTRAINT FK_SalesOrder_Customer FOREIGN
KEY REFERENCES dbo.Customer,
       OrderTime DATETIME2(3) NOT NULL,
       SalesAmount DECIMAL(18,4) NOT NULL,
       TaxRate DECIMAL(6,4) NOT NULL,
       ShipTime DATETIME2(3) NULL,
       ReceivedTime DATETIME2(3) NULL);

This table in Listing 1-1 contains only eight columns that provide information on
orders, including the customer, product details, and timing/cost. After some time and
deliberation, an analytic team creates a data warehouse–style fact table based off of this
transactional table, as seen in Listing 1-2.

Listing 1-2. An Example of an Analytic Sales Order Table

CREATE TABLE fact.SalesOrderMetrics

(      OrderDate DATE NOT NULL,
       CustomerID INT NOT NULL,
       OrderCount INT NOT NULL,
       SalesAmountTotal DECIMAL(20,4) NOT NULL,
       SalesAmountMin DECIMAL(20,4) NOT NULL,
       SalesAmountMax DECIMAL(20,4) NOT NULL,
       AverageTaxRate DECIMAL(6,4) NOT NULL,
       MinTaxRate DECIMAL(6,4) NOT NULL,
       MaxTaxRate DECIMAL(6,4) NOT NULL,
       AverageHoursFromOrdertoShip DECIMAL(6,2) NULL,
       AverageHoursFromShiptoReceive DECIMAL(6,2) NULL,
       MinimumSecondsBetweenOrders INT NULL,
       MaximumSecondsBetweenOrders INT NULL);

3
Chapter 1 Introduction to Analytic Data in a Transactional Database

Note that the analytic table in Listing 1-2 has a variety of additional metrics that are
used to summarize data elements as the source data is transformed from detail data into
daily aggregate data. Analysts could easily add many more columns if an organizational
need existed for them.
Because of these details, analytic data can easily grow and encompass millions or
billions of rows in a destination table. Similarly, an analytic table could contain far more
columns than its source transactional data. As a result, it should not be surprising that an
analytic table can potentially have many rows and/or many columns. Even a table that is
architected to not be wide (many columns) or deep (many rows) can change over time
as an organization grows and evolves. Therefore, having plans to manage larger data can
be beneficial, even if those plans are not implemented when an analytics project is first
completed.

Analytic Data Structure

Analytic data will typically be comprised of one of two kinds of data:

• Statistical metrics

• Detail data

Statistical metrics are numeric fields that can be aggregated into meaningful derived
metrics. These may be dates, durations, measurements, rates, and more. In general,
these data types use fixed-length data types, are relatively small, and have a predictable
storage footprint.
Detail data consists of text-based data types, such as VARCHAR, JSON, XML, and
other types of markup. This data is typically not used directly in metrics or statistics, but
instead is further crunched for analysis after the fact. While detail data is important, it
is not the focus of this book. If a dimension exists with a large text-based data type, it is
worth considering the normalization of that column to reduce its footprint and allow for
easier management of detail values.
For example, the table in Listing 1-3 contains a large text column.

4
Chapter 1 Introduction to Analytic Data in a Transactional Database

Listing 1-3. Example of a Table with a Wide Text Column

CREATE TABLE dbo.WebAccessLog

(      LogId INT NOT NULL,
       LogTime DATETIME2(3) NOT NULL,
       LogSource VARCHAR(250) NOT NULL,
       ErrorCode BIGINT NOT NULL);

Assuming that the data in the LogSource column is repeated often, normalizing
it into its own dimension table would save computing resources and allow for easier
analysis of distinct values or a subset of values, when needed. Listing 1-4 illustrates this
table after LogSource is normalized.

Listing 1-4. Normalization of a Text Column into a Lookup Table

CREATE TABLE dbo.LogSource

( LogSourceId SMALLINT NOT NULL IDENTITY(1,1) CONSTRAINT PK_LogSource
PRIMARY KEY CLUSTERED, LogSource VARCHAR(250) NOT NULL);

CREATE TABLE dbo.WebAccessLog

(      LogId INT NOT NULL,
       LogTime DATETIME2(3) NOT NULL,
       LogSourceId SMALLINT NOT NULL,
       ErrorCode BIGINT NOT NULL);

The LogSource column, which had been represented with a 250 character string, now
uses a SMALLINT to reference the lookup table LogSource instead. When normalizing
a column, ensure that the data type for the lookup key is neither too small (and could
run out of values) nor too large (wasting space). Normalizing dimensions is not always a
necessary action, but it can be beneficial when dimension values are often repeated and
the need to reference them is frequent.

Analytic Data Sources

Where can analytic data be stored, and how can it be known that this decision is being
made effectively? With a brief review of analytic data out of the way, this question of
where to store analytic data can be explored in a bit more detail.

5
Chapter 1 Introduction to Analytic Data in a Transactional Database

Data Warehouse
A data warehouse is a classic repository for OLAP data structures. Decades of experience
have gone into the development of exceptionally effective data warehousing software
with nearly every major data platform having its own proprietary warehousing solution.
Solutions such as Microsoft’s Azure Synapse, Amazon RedShift, and Google
BigQuery are examples of data warehousing solutions that are served up with the intent
of being used alongside other data storage components that these organizations offer.
The upside of using these technologies is that they are prebuilt, ready to use, and are
optimized for analytic data. The downside of these systems is that they involve learning
a new architecture, new scripting languages, and gaining enough expertise to maintain
them over time. Larger organizations can often afford to specialize in this regard, but
even then, the desire to increase complexity is sometimes a reason to not start with one
of these solutions.

Unstructured Data
Not all data should go into a SQL database or even a structured data source. Data that
is heavily comprised of text and markup is not easy to search within the confines of a
transactional database, even with features such as SQL Server’s Full-Text Search enabled.
When the data being worked with is dollars, time, or metrics, there is great flexibility
in where that data can effectively be stored, but large text and files make more sense
to be stored in an unstructured data repository, such as a data lake underpinned by
Hadoop, Hive, or a similar technology offered by a given software vendor.
If a data storage solution is needed and the source data is large and text based,
consider destinations outside of the confines of a transactional database. While this book
is not intended to explore these solutions, knowing when they should be considered is
helpful to avoid architecting around the wrong technology for that data.

Third-Party Analytics Software

There are many analytic software solutions available for use both on premises and in the
cloud. Even creating a list of these solutions would take hours, consume many pages,
and likely still be incomplete.

6
Chapter 1 Introduction to Analytic Data in a Transactional Database

Like with a data warehouse solution, the primary benefit of a third-party analytic
solution is that it is built for this purpose and can be used with a relatively minimal
amount of effort. Unlike a data warehouse solution, analytics software varies greatly in
its structure, interface, and usage. An organization’s data analytic needs will dictate the
type of solution needed, and some analytics software will meet those needs while some
will not.
The downsides of using a third-party vendor are cost and vendor lock-in. Costs
typically increase as data volume and/or usage increases. Therefore, an analytics
project may begin inexpensively, but end up being quite costly when the data is ten
or a hundred (or a million?!) times larger than its initial size. When architecting any
data solution, be certain to determine cost based on current and future needs. This
ensures that there are no surprises when data and its usage inevitably grow by orders of
magnitude from where they begin.
Vendor lock-in may also pose a challenge. Moving data into an analytics solution
is easy, but is there a well-documented and convenient way to extract that data in
the future? If not, then moving away from that solution may become costly and
time-consuming. If those costs exceed the costs of remaining on that platform, then an
organization may have to begrudgingly choose to maintain its data there indefinitely.
For those that are comfortable and familiar with a given data analytics solution, this is
likely not an issue, though.

OLAP Data in an OLTP Table in a Transactional Database

Storing analytic data in OLTP tables is likely a trap. OLTP tables are intended for workloads
in which small ranges of rows are frequently written and read. In these scenarios, many
columns are typically requested from those rows. These tables are designed to manage
concurrency based on the idea that a given transaction will involve few rows.
OLTP tables are counter to analytic data, which often involves queries against many
rows in aggregate, but usually not as many columns.
For small tables (less than a few million rows), an OLTP table is adequate for analytic
data. For larger tables that contain millions or billions of rows, these tables will have
difficulty in serving up data across large ranges. They will likely consume more space
and drive slower queries against their data than any other analytic solutions discussed
here. This is because they are optimized for small range lookups (few rows, many
columns) and are not architected to service the vast row counts and aggregates required
for analytics.
7
Chapter 1 Introduction to Analytic Data in a Transactional Database

Reserve OLTP tables for small analytic data sets where scanning most (or all) of a
table is not prohibitively expensive. Alternatively, they can be used for lookup tables or
other small dimensions. For a previous example in Listing 1-3, the LogSource table used
a SMALLINT to number its rows. With a primary key that can contain no more than
65,535 values (-32,767 through 32,767), this table is destined to be relatively small. It is
compact enough that indexing the text column would not be prohibitively expensive,
if needed.

LAP Data in an Analytic Table in a

O
Transactional Database
The final option for storing analytic data is an OLAP table within a transactional
database. This is the focus of the remainder of this book and will provide an alternative
to the solutions presented thus far.
An OLAP table is a table structure that is designed specifically to service queries that
typically access a large volume of rows in aggregate, but also do not require returning all
of the columns at any one given time. An OLAP table is expected to grow to be millions
or billions of rows and can easily service queries that crunch millions of rows at a time
with ease.
There are many benefits to storing analytic data in an OLAP table within SQL Server,
including the following:

• Can use existing technology and licenses to maintain analytic data.

• Far smaller learning curve to implement and use.

• Data is maintained completely by its owners.

• Data can be stored in close proximity to its transactional data source.

These are not trivial considerations and can make the effort to implement an analytic
solution fast, inexpensive, and easy to revise, if needed. In addition, performance can be
measured and optimized as needed to ensure that the end users of analytic data enjoy
acceptable access speeds. In addition, the processes that load data into analytic tables
can be fine-tuned to be exceptionally fast.
Given their simplicity and ease of implementation, it is worthwhile to take the
time and resources to create a proof of concept for storing analytic data in SQL Server,
specifically using columnstore indexes.

8
Chapter 1 Introduction to Analytic Data in a Transactional Database

Analytic data can be large, consume immense resources, and be challenging to move
when change is needed. Choosing the correct target data structure for analytic data
and making the best possible initial decision is critical to the long-term success of an
analytics project.
This book will explore in detail the nature of analytic workloads and why
columnstore indexes in SQL Server can provide a cost-effective and performant solution.

9
CHAPTER 2

Transactional vs. Analytic

Workloads
Transactional and analytic data are fundamentally different in terms of how they are
written, read, stored, and managed. To achieve optimal performance of either type
of data requires architecting data structures and software that cater to each of these
paradigms.
Choosing the optimal data structure is critical in large, busy production systems as
the cost to rebuild an existing system is significantly higher than to architect it effectively
from the start. This chapter explores the differences in these workloads and will allow
for the effective introduction of columnstore indexes as the optimal solution for a wide
variety of analytic data storage needs.

T ransactional Data
The data that is readily created, altered, deleted, and read by applications is transactional
data. This data may not always be stored in structured databases or adhere to the mantra
of ACID transactions (Atomic, Consistent, Isolated, Durable), but for this discussion, it
will all be considered transactional data.
Consider an order system that handles product orders from a website and a popular
mobile app. This is decidedly a transactional system. Many people access the site
regularly and create orders, check their status, and await the arrival of what they’ve
ordered. The WideWorldImporters sample database contains an Orders table that is
representative of what would be seen in a typical order system, as seen in Figure 2-1.

11
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_2
Chapter 2 Transactional vs. Analytic Workloads

Figure 2-1. Example of an order table used by a busy order processing system

Note that there are many columns, each of which represents a detail of an order,
such as the salesperson, customer, contact, order date, and purchase order number.
Regardless of how large this table became over time, the average user would usually
not want to view more than a handful of rows at a time. The quantity of data accessed
is typically small as users will rarely view more than current or recent orders. If they do,
pre-existing filters and links would be built to ensure that data can be served up quickly.
Oftentimes, many columns are returned to the application at one time. When
viewing an order, it is far faster and efficient to return any column that may be needed
rather than be forced to return to the table multiple times to retrieve more columns.
Because users are interacting directly with an application, there is an expectation
that performance will be as fast as possible. Waiting a second or two for an order to
process would be seen as acceptable, whereas waiting 1 or 2 minutes would result in
immediate complaints and lost business. Similarly, because many users are accessing
the site at one time, contention is expected to be high. Since many people will be
accessing their orders at one time, the tables storing that data need to be capable of fast
reads and writes, even when thousands (or millions) of orders are being accessed at
one time.
In general, transactional queries tend to be simple. For example, a user that is
retrieving information on a current order may do so via the following query:

SELECT
       Orders.OrderID,
       Customers.CustomerName,
       Orders.OrderDate,
       Orders.ExpectedDeliveryDate,
       Orders.CustomerPurchaseOrderNumber,
       OrderLines.*

12
Chapter 2 Transactional vs. Analytic Workloads

FROM Sales.Orders
INNER JOIN Sales.Customers
ON Customers.CustomerID = Orders.CustomerID
INNER JOIN Sales.OrderLines
ON OrderLines.OrderID = Orders.OrderID
WHERE Customers.CustomerID = 10
AND Orders.OrderDate >= '5/20/2016'
AND Orders.OrderDate < '5/27/2016';

This is a relatively simple query that pulls 5 specific columns, as well as all 12
columns of line-item detail from 3 tables. Figure 2-2 shows the results of this query.

Figure 2-2. Results of a sample transactional order query

Note that only two rows are returned via a filter on both customer and date. It would
be unusual for a user of this system to request data on hundreds or thousands of orders
at once. Viewing so much data at one time would be challenging, and it is unlikely that
the user interface is built to allow so much data to be retrieved at one time anyway.
Any data that is served up by a transactional system needs to be accurate at runtime.
Someone checking their bank balance needs to get an accurate number without
exception. This means that data integrity needs to be enforced at runtime either by the
database, the application, or a combination of both. Inaccurate data in a transactional
system could have disastrous results if the application in question, for example, belongs
to a hospital, the military, or meteorologists.
Similarly, if a transactional operation fails, that failure needs to be addressed
immediately so that that user’s data is not left in an uncertain state. If a user renews their
passport online, but the payment never correctly processes, their application may not
be handled as it should be. Worse, there may not be an easy way for the user to fix it.
Therefore, the ability of that system to properly handle each step of that process together
as a unit is very important. The use of ACID for transactional data is heavily dependent
on the data and its use. A bank cannot afford an inconsistent transaction, but a social
media site probably can. Similarly, data that is eventually consistent will be perfectly
acceptable in some organizational models, whereas it would be destructive in others.

13
Chapter 2 Transactional vs. Analytic Workloads

OLTP data is often normalized to assist in ensuring relational integrity and the
ability to generate lists of lookup values efficiently. Normalization can save space and
memory in transactional applications as more verbose text columns are replaced with
numeric lookups. It also allows for easy lookups within the application by querying the
normalized table directly.
In summary, transactional data generally shares these characteristics:

• Often the target of INSERT, UPDATE, and DELETE operations.

• Many columns are returned at one time, for use by the application.

• Volume of data accessed by a single user at one time is

typically small.

• Transaction speed is paramount.

• High levels of contention are expected.

• Data must be accurate at runtime, when needed, without exception.

• Accessed by application end users.

• Often normalized.

While these characteristics are generally true of transactional data, analytic data is
quite different and is worth exploring further in detail.

Analytic Data
While transactional data is found in live/busy systems processing orders or bank
operations, analytic data is found further downstream. This data is most often associated
with reporting, visualization, and analytics. It is what powers business intelligence, data
science, and the decision making that drives organizations worldwide.
OLAP systems are fundamentally different from transactional systems. To ensure
optimal speed and reliability, each needs to be architected, implemented, and managed
in distinctly different ways.
Analytics typically require queries that access large batches of data at one time,
but will need far fewer columns to perform the calculations needed to drive results.
For example, a common financial request may be to compare revenue for the current

14
Chapter 2 Transactional vs. Analytic Workloads

quarter with revenue from the previous quarter. The calculations needed to drive this
request would need to access all revenue data for each quarter or have that data
pre-crunched in a convenient location.
The following query is an example of an analytic data request that seeks to
understand order counts and totals over time:

SELECT
       Date.[Calendar Year] AS Order_Year,
       Date.[Calendar Month Number] AS Order_Month,
       COUNT(*) AS Order_Count,
       SUM(Quantity) AS Quantity_Total,
       SUM([Total Excluding Tax]) AS [Total Excluding Tax]
FROM Fact.[Order]
INNER JOIN Dimension.Date
ON Date.Date = [Order].[Order Date Key]
WHERE [Order].[Order Date Key] >= '1/1/2016'
AND [Order].[Order Date Key] < '1/1/2017'
GROUP BY Date.[Calendar Year], Date.[Calendar Month Number]
ORDER BY Date.[Calendar Year], Date.[Calendar Month Number];

Note that only a handful of columns are requested, but that almost 30k orders were
queried to return the results shown in Figure 2-3.

Figure 2-3. Results of a sample analytic query

A larger OLAP system could easily process millions of rows as a part of a single
request with similar-looking results. An effective analytic data store needs to be able
to scale to support millions (or billions) of rows in any given table without consuming
excessive system resources or queries taking too long to execute.

15
Chapter 2 Transactional vs. Analytic Workloads

Unlike transactional systems where data is written in small batches from many
sources, analytic data is typically created in larger bulk operations from a limited set
of organized data load processes. Inserting thousands (or millions) of rows at one
time needs to be fast and efficient as that will be the norm. The diagram in Figure 2-4
illustrates the differences between how data is written in each type of system.

Figure 2-4. Summary of OLAP vs. OLTP data access patterns

It is important to note that not all OLTP and OLAP systems behave precisely like this,
but will generally follow these usage patterns.
Because analytic data is typically used for reporting, analytics, and data science, the
need for real-time results is less important. Many analytic processes are asynchronous
and not directly monitored by the user that will consume the resulting data. Because
OLAP data changes less often, complex data processing can occur prior to data usage,
ensuring that there is no need to wait too long for results, if the need for interactive
reporting is important.
Similarly, because the number of data sources, users, and apps accessing the data
is smaller, contention is far less of a challenge. A query may require reading millions of
rows from a table to return its results, but the quantity of queries accessing that table will
be orders of magnitude less than a busy transactional table.

16
Chapter 2 Transactional vs. Analytic Workloads

Because OLAP data is updated infrequently, the need to maintain relational

integrity via constraints is lower. The set of new data that is loaded can be independently
validated and confirmed to be accurate prior to its insertion into an analytic table.
Foreign keys and constraints may be still used as conveniences for joining or reporting,
but their use will depend on their impact on load speeds. The decision on how to
validate analytic data depends on each use case, the nature of its source transactional
data, and the amount of change that can occur in the OLAP data outside of organized
data load processes.
Similarly, architects of OLAP data load processes can take liberties with how they
handle failures. In transactional applications, failures need to be immediately retried or
rolled back. There can be no middle ground that could result in bad or inconsistent data.
A failed analytic load process can be resolved with a more brute-force approach such
as this:

1. Troubleshoot and fix the broken load process.

2. Identify and remove all newly loaded data.

3. Rerun the load process.

There is likely no need to preserve any of the partially loaded data, nor is the removal
of partially loaded data complex, as it is typically time-stamped. Many OLAP data load
processes will include steps that automatically remove any data previously loaded for
the same time period. This allows a failed process to be rerun without the prerequisite
cleanup step.
The consumers of analytic data are also quite different from those that use
transactional applications. Whereas an OLTP application may have a wide range of end
users that regularly interact directly with the software, analytic data consumers tend to
be analysts, domain experts, and business leaders. Domain experts are people that are
close enough to the data to understand its meaning and usage. For example, a network
administrator would be a domain expert for network traffic data, whereas a chief
financial officer would be a domain expert for quarterly earnings data. Domain experts
may have a clear understanding of data meaning, but they are less likely to be familiar
with how that data is stored and processed.
Unlike OLTP data, analytic data is often denormalized. Lookup tables may be
maintained for reporting purposes, but not used exclusively as lookups to fact tables.
Normalization does not always save significant storage or memory resources in OLAP
tables. Equally important, joining lookup tables during data load processes can greatly

17
Chapter 2 Transactional vs. Analytic Workloads

increase their latency, resulting in longer delays before new analytic data becomes
available. Most compression algorithms used in mainstream analytic data structures
utilize dictionary lookups to “normalize” data when it is physically stored. This means
that repeated values take up far less space than in a transactional table, regardless of
the column’s specific width. Despite that, dictionaries are limited in size and wider data
types will pose compression challenges, which will be tackled in Chapter 5.
Analytic data can be summarized as follows:

• Written by large, infrequent, and controlled data load processes.

• Rows are read in large batches for use in analytics.

• Less columns are typically read in each operation.

• Low levels of contention are expected.

• Data integrity can be enforced via validation, rather than

constraints.

• Accessed by business leaders, analysts, and domain experts.

• Often denormalized.

The characteristics of analytic data are quite different from those of transactional
data. These different characteristics tend to lead toward using different systems to
store each.

The Need for Two Systems

Because the workloads of transactional and analytic systems are so vastly different, the
need for unique data structures to maintain them becomes more apparent the larger
their data gets.
While a thousand rows of data can be quickly accessed in any structure, a billion
rows would be cumbersome unless its data structure was built to accommodate the
types of workloads it is expected to service.
Many applications attempt to mix analytic and transactional workloads to maintain
convenience and avoid the creation of new systems. This works until the various
differences between these types of data begin to clash with each other. For example, a
common pitfall in mixing OLAP and OLTP workloads occurs when large, slow analytic

18
Chapter 2 Transactional vs. Analytic Workloads

queries cause locking and blocking against objects that application users are trying
to write to. If a salesperson running an annual report is blocking the users trying to
purchase new products, then the result will be lost sales and disgruntled users.
Similarly, those large reports may consume vast amounts of memory, forcing
common transactional data out of the buffer pool in favor of reporting data. This will
further exacerbate slowness for application users.
In general, mixing analytic and transactional workloads may not always be
avoidable, but considering the ramifications of this architectural decision up front can
help in predicting (and planning for) problems before they become costly.
Many other considerations will vary between OLTP and OLAP systems, such as

• Maintenance

• High availability

• Disaster recovery

• SQL Server licensing

• Hardware resources

The architect of an application will need to ensure that each of these topics is visited
separately for both transactional and analytic workloads. While it is possible that similar
systems can be used when data sizes are small, it is unlikely that those same systems
will scale indefinitely. Eventually, data becomes large enough that unique solutions are
required to differentiate each type of workload.
The nuance of these decisions can travel deeper. For example, a small analytic
database would likely be architected differently than a large data repository. Therefore,
it’s important to consider size, scale, data growth, and workload together when
architecting new data structures or when revisiting existing database objects.

Building Better Analytic Data Structures

The discussion thus far makes it clear that the tables and structures used to support
analytic data need to be distinct from what is used for transactional workloads. In this
chapter, many details were outlined that differentiate OLTP and OLAP applications, how
they read and write data, and their typical usage patterns.

19
Chapter 2 Transactional vs. Analytic Workloads

Taking this information into account allows for a vision to be created of where
analytic data should be stored. The remainder of this book will focus on how
columnstore indexes in SQL Server can be used to effectively meet the needs of large
analytic tables and workloads. This technology is mature and can be used to effectively
manage a wide variety of analytic use cases, ensuring that it provides an ideal solution
to many of the analytics challenges that architects and administrators face on a
regular basis.

20
CHAPTER 3

What Are Columnstore

Indexes?
Thus far, a compelling case has been presented to separate OLTP and OLAP workloads
into unique data structures and optimize each for their respective use cases. Columnstore
indexes are a SQL Server feature that provides native support for large analytic data. This
chapter will dive into what they are and why they are an effective solution to analytic data
challenges.

The Limits of Transactional Data Storage

Tables that contain millions or billions of rows of analytic data do not scale well when
implemented as classic transactional tables. To fully understand why this is the case, it
is helpful to compare how data is stored in both transactional and analytic tables. The
following example table contains five columns. Figure 3-1 shows how data is stored in an
OLTP (rowstore) version of this table.

Figure 3-1. Illustration of data storage in a clustered rowstore index

In a clustered rowstore index, each row is stored on pages sequentially, one row after
another in the order prescribed by the clustered index. If many columns from a single
row or small subset of rows are required by a query, then this is an exceptionally efficient

21
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_3
Chapter 3 What Are Columnstore Indexes?

storage structure. For example, retrieving all columns from row 1 through row -5 would
require low effort as the data is contiguous and ordered. Figure 3-2 highlights the data
needed to satisfy this query.

Figure 3-2. How an OLTP query retrieving five rows would read a sample table

A table with sequential columns models how typical transactional queries operate
for a query such as this:

SELECT
       OrderID, --  An identity/primary key column
       CustomerID,
       SalespersonPersonID,
       ContactPersonID,
       OrderDate
FROM Sales.Orders
WHERE OrderId = 289;

Transactional queries write single rows or small ranges of rows that often correlate
to a lookup on a single index. Here, a numeric identity is used to filter out five rows of
interest from a larger table.
The basic unit of SQL Server storage is the page. A page is comprised of 8 kilobytes
of data. Pages contain all data and index storage for a given table. If a single row is
required for a query, then the entire page it is stored on will be read into memory, even
if the remaining data on the page is not needed. Therefore, transactional queries rely
on clustered or nonclustered indexes to ensure that index seeks can return data in an
ordered fashion, such as the example in Figure 3-2. In this scenario, reading those five
rows will not require much more effort, even if the table grows to be significantly larger.
Analytic queries, though, are quite different. They often aggregate a select few
columns, but do so across many rows. Consider a typical OLAP query against the
example table where a single column is aggregated across a large portion of the table that
happens to include all of the rows shown here. Figure 3-3 shows how this would look.

22
Chapter 3 What Are Columnstore Indexes?

Figure 3-3. An analytic query aggregating a single column from a sample table

The physical layout of rows shows that a query requiring only column 2 needs to also
read the adjacent columns, even if not needed. The following is an example of a query
that accesses a large amount of data in a table, but only aggregates a single column.

SELECT
SUM(Quantity) AS Total_Quantity
FROM Sales.OrderLines
WHERE OrderID >= 1
AND OrderID < 10000;

Even though only a single column was summed, any page that contains values for
that column will need to be read into memory. Even a covering nonclustered index
would still necessitate reading data for each row, though the number of pages read can
be reduced in that fashion.
Now consider a larger table with 50 million rows that is stored on 250,000 pages
using a clustered rowstore index. The transactional query demonstrated in Figure 3-2
would still only read 5 rows that are stored consecutively and could therefore ignore
most of the remaining 249,995 rows and the pages they are stored on. Additional rows
would be read that happened to be stored on the same page(s) as those five rows, but the
added burden of that data is measured in kilobytes and is comparatively insignificant.
The analytic query presented in Figure 3-3 aggregates only a single column, but does
so across many rows. If the requested order data spans one-fourth of the table, then
processing this query in the transactional table would force SQL Server to read about
one-fourth of the pages in the table, since the values for column 2 are dispersed within
each row throughout the table. Regardless of the filter used that reduces the row count
needed for the query, every page in the range specified by the filter would need to be read.
As an analytic table continues to grow and spans billions of rows and/or terabytes of
storage (or more), the ability to read large swaths of its data becomes too slow and resource
intensive to be realistic. A better solution is needed that allows analytic queries to read
column data without the rest of the underlying columns being brought along for the ride.

23
Chapter 3 What Are Columnstore Indexes?

Introducing Columnstore Indexes

The technology needed to solve this problem has existed in some form for a long time.
Applications such as SQL Server Analysis Services and PowerPivot have used columnar
data storage formats for years, but their implementation was hidden from view of
the user.
Columnstore indexes allow for data to be stored in tables in an optimal analytic
format that provides many valuable benefits, such as

• Ability to scale to any table size

• Increase query speeds by 10–100 times

• Exceptional compression ratios, saving storage and memory

• Natively supported by SQL Server

• Can take advantage of bulk loading for fast write speeds

While this may sound like a sales pitch, there are no exaggerations here. Comparing
a large analytic data set stored in a classic rowstore clustered index vs. a columnstore
indexed table yields vast differences in storage and performance that will be
demonstrated and quantified throughout the remainder of this book.
Note that all demonstrations of columnstore indexes in this book are tested on
SQL Server 2019. Those running an earlier version should test thoroughly before
implementing any suggestions in this book as the features available may be different.
Figure 3-4 illustrates how data stored in a columnstore index is stored using the same
table from Figure 3-1.

Figure 3-4. Illustration of data storage in a clustered columnstore index

24
Chapter 3 What Are Columnstore Indexes?

Note the critical difference: Data is ordered by column rather than by row. Each
value for column 1 is stored together in a single structure, whereas each other column
is stored in its own sets of pages. In the rowstore table shown in Figure 3-3, a single
column was aggregated, but every page had to be read in order to return data for that
one column.
Figure 3-5 shows the same query against an analytic columnstore indexed table.

Figure 3-5. Aggregating a single column against a columnstore index

Because data is grouped physically by column, it is no longer necessary to read

the entire table to return a single aggregated column. The other four columns can be
ignored. This greatly reduces reads against storage systems and reduces memory usage
as less pages need to be read into memory. For each new row that is inserted into this
table, only a single additional value for the second column needs to be read.
A significant benefit of this data structure is that compression becomes more
effective. A set of values for a single column will tend to have more repeated values
that will more easily compress, whereas values from different columns are less likely to
overlap and compress well. Better compression means more data can fit on pages, saving
additional storage and memory.
An important clarification on columnstore indexes is that they are not “just another
index.” They are not comparable to nonclustered rowstore indexes, XML indexes, spatial
indexes, memory-optimized indexes, or other types of indexing included in SQL Server.
They comprise a unique architecture that provides benefits well beyond what a typical
index can deliver.
Note that throughout this book, when not otherwise specified, all demonstrations
and discussion will reference clustered columnstore indexes. Chapter 11 will delve into
non-clustered columnstore indexes in more detail.

25
Chapter 3 What Are Columnstore Indexes?

Benefits of Columnstore Indexes in SQL Server

There are many reasons why columnstore indexes can be an ideal solution for storing
large analytic data. They cross a variety of areas from cost to convenience to speed and
illustrate how a seemingly simple feature can provide exceptional value with a low time
and resource cost to implement. This is not a complete list of the benefits of columnstore
indexes, but highlights the keys to performance and efficiency that make them attractive
for use with analytic data. The remainder of this book will dive into greater detail about
the benefits and optimal use cases for columnstore indexes.

Native Analytic Data in SQL Server

One of the greatest benefits of columnstore indexes is that they store analytic data
directly in SQL Server without any added licensing or configuration changes. Similarly,
no hardware or software changes are required to begin testing or implementing
columnstore indexes.
A rowstore table containing analytic data could be converted to use a columnstore
index, with the entire process taking place on the same SQL Server database instance
or even as an in-place index swap. This can allow for easy testing, validation, and
implementation, when ready.
Because columnstore indexes are a feature that is core to analytic data storage in
SQL Server, it is updated with each new version. Each update provides new functionality
and ways in which reading or writing columnstore indexes can become faster and more
efficient. The following document from Microsoft outlines in detail the columnstore
index features available in each version of SQL Server since their inception:

https://docs.microsoft.com/en-us/sql/relational-databases/indexes/
columnstore-indexes-what-s-new

The list is quite extensive and shows how columnstore indexes have evolved from an
inflexible read-only structure into one rich in features and optimizations.
Storing data natively in SQL Server means there is no need for third-party products,
no costly migrations, and no need to configure new hardware and software. The
time required to implement them is less guided by technology and more by typical
development and quality assurance needs. When building a project plan for the storage

26
Chapter 3 What Are Columnstore Indexes?

and maintenance of analytic data, factoring in these resource costs can help in making
an accurate,
fact-based decision. The following is a summary of these considerations:

• Need to purchase/license new hardware to support analytic data

• Cost of software licenses for analytics software and any supporting

software

• Cost of computing resources, whether on premises or in the cloud

• Time required to architect, develop test, and implement analytic data

solution

• Resources needed to educate people on how to use the new solution

Quantifying each of these factors can assist in comparing and contrasting different
solutions and will generally provide favorable results for an organization that already
uses SQL Server for its transactional data storage.

Scalability
Given the rapid rate in which analytic data can grow, any data structures used to store it
must be capable of efficiently servicing OLAP workloads, even when data depth or width
increases unexpectedly fast over time.
The ability for an analytic data storage solution to scale has many important benefits,
including
• Ensure high performance, even during periods of rapid data growth

• Avoid costly migrations to new technologies

• Avoid disruptions to reporting/analytic services

• Reduce maintenance and downtime

• Reduce the need for hardware upgrades

For an OLAP solution to be effective, it needs to be fast and efficient with one
thousand rows, one million rows, one billion rows, or more. If a system is destined to
become inefficient when it gets large, then it is also destined to fail.

27
Chapter 3 What Are Columnstore Indexes?

The growth of analytic data is a metric worthy of measuring and revisiting

periodically to fully understand its long-term resource requirements. OLAP data rarely
gets smaller, and its growth rate rarely decreases. This growth is ultimately tied to two
measurements for any given organization:

• Growth

• Complexity

As an organization grows and serves more customers, those customers will generate
more data. Similarly, organizational growth almost always leads to both technical and
nontechnical processes becoming more complex. This complexity may manifest itself in
increased software features, more processes that require tracking, or requests for more
types of data to maintain and crunch over time.
The rate of growth of data can be summarized as follows:

1.00 * N * G * C

In this representation, the letters denote the following:

• N = Natural Growth: This is the growth of data that occurs currently.

If nothing else changes and no external influences adjust the data
growth rate, then this would be the only factor needed to predict
future data growth. This can be measured in rows per unit time and
also as an increase in physical data size as average bytes per row
times rows per unit time.

• G = Data Growth: Natural growth provides a baseline of how much

data will be added to a table or database over time. This growth is
rarely linear, though. New customers, increased sample frequency,
and other factors will account for additional growth above and
beyond the current baseline. This can also be measured in rows per
unit time or as physical data usage per unit time.

• C = Increased Complexity: This accounts for more columns being

added to tables as well as new tables being created and populated.
This also increases the amount of data that needs to flow (via more
entities) from transactional systems into the target analytic system.
This is challenging to measure, but can be estimated as a measure
of new metrics over time. That is, on average how many new metrics

28
Chapter 3 What Are Columnstore Indexes?

per unit time will be added to an analytic data source. This can be
approximated in storage units as soon as the data types for new
metrics are known.

Each of these factors is defined here as linear in nature, but may not be linear
in actuality. Therefore, they should be revisited regularly to ensure that unexpected
changes in growth are accounted for when predicting future data size. Linear
approximations per unit time can be used to approximate nonlinear growth so long as
those approximations remain updated with current and future trends.
To provide a sample of the preceding formula, consider an analytic table that
contains 100,000,000 rows, currently grows by 250,000 rows per day, and is expected to
see its growth accelerate by an additional 25% per year, but not see any new dimensions
added in the foreseeable future. The estimated row count in 1 year would be given by

1.00 * N * G * C * Row Count = 1.00 * 1.9125* 1.25 * 1.00 *

100,000,000 = 239,062,500 rows for the upcoming year.

1.00 * N * G * C * Row Count = 1.00 * 1.9125* 1.25 * 1.00 *

239,062,500 = 571,508,789 rows for the 2nd year.

1.00 * N * G * C * Row Count = 1.00 * 1.9125* 1.25 * 1.00 *

571,508,789 = 1,366,263,198 rows for the 3rd year.

Note how quickly the added acceleration grew this data – from 100 million rows to
over 1.3 billion rows in 3 years! 1.9125 was calculated as the annual growth factor by
multiplying 250,000 rows per day by 365 days per year.
To summarize, any technology that manages analytic data needs to be capable of
efficient data access given that typical data growth can cause data sizes to balloon far
faster than conventional wisdom might suggest. The architecture of columnstore indexes
will be discussed in detail in Chapter 4 and will lay the foundation for explaining why
they can scale so effectively for rapidly growing data.

E xceptional Compression
Any analytic data store needs to take full advantage of compression to make its data
as compact as possible. Columnstore indexes use multiple compression algorithms to
achieve impressively high compression ratios. This can result in data that is 10–100 times
smaller than it would be in an uncompressed table.

29
Chapter 3 What Are Columnstore Indexes?

Compressing data by column rather than row is inherently more efficient for a
number of reasons:

• A single column contains values of the same data type, improving the
chances of values being repeated.

• Dimensions often have many repeated values and will compress

exceptionally well.

• Data is typically created sequentially, ordered by time. Rows that are

in close proximity to other rows from similar time periods will tend to
contain more similar data and compress better than with data from
much earlier or later in time.

Chapter 5 dives into more detail on columnstore index compression and

demonstrates how effective these compression algorithms are and why they are crucial
to optimal analytic query performance.

Faster Analytic Reads

First and foremost, columnstore indexes are built to provide highly performant analytic
query speeds. Whether accessing data from structured stored procedures and code or
from ad hoc analytics and visualization, the compact and segmented columnstore index
structure allows for the requested data to be returned quickly and efficiently. This is true
even when large row counts are scanned.

Faster Data Loads

When data is written millions of rows at a time, processes need to exist that allow for
that data to be written as quickly as possible. A fully logged process optimized for
transactional data will not perform adequately for data of this size.
Columnstore indexes can make full use of the SQL Server bulk load API and can
achieve write speeds far faster than other processes available in SQL Server, and in many
other products. Data loads can execute faster, improving OLAP availability and allowing
new data to appear more quickly. In addition, the impact of data loads on system
resources such as the transaction log and on backups is minimized, ensuring that large
volumes of data can be written in short spans of time without the pitfalls of doing so on a
fully logged OLTP table.

30
Chapter 3 What Are Columnstore Indexes?

Chapter 8 will fully explore how data is written to columnstore indexes, including
performance measurements, resource consumption details, and best practices for
loading data as efficiently as possible.
Analytic data requires a versatile, scalable, and performant solution. Columnstore
indexes provide an ideal data structure to create and analyze analytic data, and that can
manage data growth over time with ease.
The remainder of this book will discuss columnstore indexes in exhaustive detail,
providing architectural details, demonstrations, best practices, and tools that can
improve their use.

31
CHAPTER 4

Columnstore Index
Architecture
A solid understanding of the architecture of columnstore indexes is necessary to make
optimal use of them. Best practices, query patterns, maintenance, and troubleshooting
are all based on the internal structure of columnstore indexes. This chapter will focus on
these architectural components, providing the foundation for the rest of this book.

S
ample Data
To demonstrate the topics presented throughout this book, a sample data set will be
created based on the Fact.Sale table in the WideWorldImportersDW database.
The data set can be generated using the query in Listing 4-1.

Listing 4-1. Query Used to Generate a Data Set for Use in Columnstore
Index Testing
SELECT
       Sale.[Sale Key], Sale.[City Key], Sale.[Customer Key], Sale.[Bill
To Customer Key], Sale.[Stock Item Key], Sale.[Invoice Date Key],
       Sale.[Delivery Date Key],Sale.[Salesperson Key], Sale.[WWI
Invoice ID], Sale.Description, Sale.Package, Sale.Quantity,
Sale.[Unit Price], Sale.[Tax Rate],
       Sale.[Total Excluding Tax], Sale.[Tax Amount], Sale.Profit,
Sale.[Total Including Tax], Sale.[Total Dry Items],
       Sale.[Total Chiller Items], Sale.[Lineage Key]

33
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_4
Chapter 4 Columnstore Index Architecture

FROM Fact.Sale
CROSS JOIN
Dimension.City
WHERE City.[City Key] >= 1 AND City.[City Key] <= 110;

This generates 25,109,150 rows of data spanning invoice date ranges of 1/1/2013
through 5/31/2016. While this data is not massive, it is large enough for suitable
demonstrations without being cumbersome to those that wish to replicate it at home.
This data set will be reused in future chapters, being placed into a number of test tables
to illustrate a wide variety of topics related to columnstore indexes, OLAP performance,
and database architecture.
For this chapter, the data will be loaded into a table without any indexes, and a
columnstore index added at the end, as seen in Listing 4-2.

Listing 4-2. Script That Creates and Populates a Columnstore Index Test Table

CREATE TABLE Fact.Sale_CCI

(      [Sale Key] [bigint] NOT NULL,
       [City Key] [int] NOT NULL,
       [Customer Key] [int] NOT NULL,
       [Bill To Customer Key] [int] NOT NULL,
       [Stock Item Key] [int] NOT NULL,
       [Invoice Date Key] [date] NOT NULL,
       [Delivery Date Key] [date] NULL,
       [Salesperson Key] [int] NOT NULL,
       [WWI Invoice ID] [int] NOT NULL,
       [Description] [nvarchar](100) NOT NULL,
       [Package] [nvarchar](50) NOT NULL,
       [Quantity] [int] NOT NULL,
       [Unit Price] [decimal](18, 2) NOT NULL,
       [Tax Rate] [decimal](18, 3) NOT NULL,
       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,

34
Chapter 4 Columnstore Index Architecture

[Total Dry Items] [int] NOT NULL,

[Total Chiller Items] [int] NOT NULL,
[Lineage Key] [int] NOT NULL);

INSERT INTO Fact.Sale_CCI

       ([Sale Key], [City Key], [Customer Key], [Bill To Customer Key],
[Stock Item Key], [Invoice Date Key], [Delivery Date Key],
        [Salesperson Key], [WWI Invoice ID], Description, Package, Quantity,
[Unit Price], [Tax Rate], [Total Excluding Tax], [Tax Amount],
        Profit, [Total Including Tax], [Total Dry Items],
        [Total Chiller Items], [Lineage Key])
SELECT
       Sale.[Sale Key], Sale.[City Key], Sale.[Customer Key], Sale.[Bill To
Customer Key], Sale.[Stock Item Key], Sale.[Invoice Date Key],
       Sale.[Delivery Date Key], Sale.[Salesperson Key], Sale.[WWI
Invoice ID], Sale.Description, Sale.Package, Sale.Quantity,
Sale.[Unit Price], Sale.[Tax Rate],
       Sale.[Total Excluding Tax], Sale.[Tax Amount], Sale.Profit,
Sale.[Total Including Tax], Sale.[Total Dry Items],
       Sale.[Total Chiller Items], Sale.[Lineage Key]
FROM fact.Sale
CROSS JOIN
Dimension.City
WHERE City.[City Key] >= 1 AND City.[City Key] <= 110;

-- Create a columnstore index on the table.

CREATE CLUSTERED COLUMNSTORE INDEX CCI_fact_sale_CCI ON fact.Sale_CCI;

The size and shape of this data can be confirmed with a simple query:

SELECT
       COUNT(*),
       MIN([Invoice Date Key]),
       MAX([Invoice Date Key])
FROM fact.Sale_CCI;

35
Chapter 4 Columnstore Index Architecture

The results are shown in Figure 4-1.

Figure 4-1. Query results showing the size and date range for a set of test data

R
owgroups and Segments
Analytic data cannot be stored in one large contiguous structure. While each column is
compressed and stored separately, the data within those columns needs to be grouped
and compressed separately. Each compressed unit is what is read into memory. If
that unit is too small, then the storage and management of a multitude of compressed
structures would be expensive, and its compression poor. Conversely, if the number of
rows in each unit is too large, then the amount of data that needs to be read into memory
to satisfy queries would also become too large.
Columnstore indexes group rows into units of 220 (1,048,576) rows that are called
rowgroups. Each column within that rowgroup is individually compressed into the
fundamental unit of a columnstore index called a segment. This structure can be
visualized using the representation in Figure 4-2.

Figure 4-2. Rowgroups and segments within a columnstore index

36
Chapter 4 Columnstore Index Architecture

Note that columnstore indexes are not built on binary tree structures like clustered
and nonclustered rowstore indexes are. Instead, each rowgroup contains a set of
compressed segments, one per column in a table. The example in Figure 4-2 is for a table
with eight columns and up to 6*220 rows, containing a total of 48 segments (one segment
per column per rowgroup). The only significant architectural convention shared
between rowstore and columnstore indexes are their use of 8KB pages to store their data.
Rowgroups are created and managed automatically as data is created in a
columnstore index. There is no cap on the number of rowgroups that can exist, nor are
there limits on the count of segments within the index. Because a rowgroup contains
up to 220 rows, a table should have many more rows than this to make optimal use of a
columnstore index. If a table has 500k rows, then it is likely to all be stored in a single
rowgroup. As a result, any query requiring data from the table would need to read
segments that contain data for all rows in the table. For a table to make effective use of
a columnstore index, it should have at least 5 million or 10 million rows so that it can
be broken into separate rowgroups of which not all need to be read each time a query is
issued against it.
The rowgroups within a columnstore index may be viewed using the dynamic
management view sys.column_store_row_groups. The query in Listing 4-3 returns
rowgroup metadata for the columnstore index created earlier in this chapter.

Listing 4-3. Script to Return Basic Rowgroup Metadata for a Columnstore Index

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       column_store_row_groups.partition_number,
       column_store_row_groups.row_group_id,
       column_store_row_groups.state_description,
              column_store_row_groups.total_rows,
       column_store_row_groups.deleted_rows,
       column_store_row_groups.size_in_bytes
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables

37
Chapter 4 Columnstore Index Architecture

ON tables.object_id = indexes.object_id
WHERE tables.name = 'Sale_CCI'
ORDER BY tables.object_id, indexes.index_id,
column_store_row_groups.row_group_id;

The results of this query can be found in Figure 4-3.

Figure 4-3. Rowgroup metadata for fact.Sale_CCI

By joining in other system views, such as sys.indexes and sys.tables, it is possible to

return additional information about the table that the columnstore index resides on.
sys.column_store_row_groups contains some additional columns of interest, such as the
compression status of the rowgroup, the row count, and its size. Note that the selection
of eight rowgroups in Figure 4-3 all contain the maximum number of rows allowed in a
rowgroup, 220.
Using rowgroup metadata allows the user to quickly measure the size of a
columnstore index and gain a basic understanding of its structure.
Within each rowgroup are segments for each column in the table. A dynamic
management view also exists that provides details about compressed segments:
sys.column_store_segments. Listing 4-4 provides a query that returns information from
this view, including joins back to the parent table and column for each segment.

Listing 4-4. Script to Return Basic Segment Metadata for a Columnstore Index

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       columns.name AS column_name,
       partitions.partition_number,

38
Chapter 4 Columnstore Index Architecture

       column_store_segments.row_count,
       column_store_segments.has_nulls,
       column_store_segments.min_data_id,
       column_store_segments.max_data_id,
       column_store_segments.on_disk_size
FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id
INNER JOIN sys.indexes
ON indexes.index_id = partitions.index_id
AND indexes.object_id = partitions.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.columns
ON tables.object_id = columns.object_id
AND column_store_segments.column_id = columns.column_id
WHERE tables.name = 'Sale_CCI'
ORDER BY columns.name, column_store_segments.segment_id;

Figure 4-4 contains a sample of the results.

Figure 4-4. Segment metadata for fact.Sale_CCI

The row count returned by the query is equal to the total number of segments in the
index, which is subsequently equal to the count of rowgroups multiplied by the number
of columns in the table. In addition to the row count contained in each segment and
its size on disk, there are details given as to whether the segment has NULLs in it, and
min_data_id/max_data_id, which provide dictionary lookups for values contained in
the segment. Details about dictionaries and how they work will be provided alongside
the discussion of compression in Chapter 5.

39
Chapter 4 Columnstore Index Architecture

The Delta Store

Writing data to highly compressed segments is a resource-intensive process. The
resources required to decompress a set of segments, write additional data to them,
compress them, and update metadata is not trivial. The effort to write a million rows to
one rowgroup is comprised of the effort needed to decompress, write, and compress
all segments in the rowgroup. The effort to write a million rows one at a time to that
rowgroup would be comparable to the same process being executed one million times.
Because of this, processes that frequently write small numbers of rows need a way
to manage those writes so that a columnstore index is not consuming all of a server’s
resources decompressing, writing, and re-creating segments. The delta store is a set of
clustered rowstore indexes that temporarily store small writes alongside the columnstore
index. Each delta rowgroup is one of these clustered indexes. Changes accumulate in the
delta store until its row count reaches a threshold (220 rows per delta rowgroup) and is
pushed all at once into the columnstore index.
The delta store is maintained automatically by SQL Server. An asynchronous
background process called the tuple mover manages the movement of data from delta
store into columnstore index. While operators can run maintenance scripts to influence
the behavior of the delta store, doing so is not necessary for using columnstore indexes.
More information on columnstore maintenance can be found in Chapter 14.
When queries are executed against a columnstore index, the contents of the delta
store are read alongside the compressed segments that are needed. While the delta store
is comprised of classic rowstore indexes, its size is generally very small compared to
the compressed portion of the columnstore index; therefore, reading it is generally not
detrimental to performance. The benefit of the delta store is that it greatly reduces the
computing resources needed to write smaller batches of data to a columnstore index and
ensures that frequent smaller writes do not become unscalable bottlenecks during a data
load, maintenance, or software release process.

40
Chapter 4 Columnstore Index Architecture

The basic flow of data can be shown in Figure 4-5.

Figure 4-5. Flow of data from the delta store into a columnstore index

Note that the delta store is not used to manage all INSERT operations against a
columnstore index. Chapter 8 discusses in detail how bulk load processes are used to
greatly speed up larger INSERT operations.

The Delete Bitmap

Deletion of data from a columnstore index poses similar challenges to those solved for
inserts by the delta store. Deleting a single row would require a compressed segment per
column to be decompressed, modified, and recompressed. This is inherently expensive
and becomes prohibitively costly when the number of rowgroups affected increases.
SQL Server needs to be able to delete data from a columnstore index quickly and
without having a prolonged negative impact on server performance. To accomplish this,
data is not physically deleted from a columnstore index when a DELETE statement is
executed. Instead, the deleted status of the affected rows is written to a structure called
the delete bitmap.
Only one delete bitmap can exist per partition within a columnstore index, and
its purpose is to track which rows are deleted (if any). When rows are deleted from
a columnstore index, the delete bitmap is updated to indicate the deletion. The
corresponding rows within the columnstore index are not modified in any way. These
soft deletes allow DELETE operations to execute quickly and not become burdensome to
data loads or maintenance processes.
When a query reads data from a columnstore index, the contents of the delete
bitmap are also read and any rows that are flagged as deleted will be omitted from the
results. The delete bitmap is also a clustered rowstore index that exists side by side with

41
Chapter 4 Columnstore Index Architecture

the columnstore index. The biggest upside of having a delete bitmap is that DELETE
operations can execute exceptionally fast as there is no need to decompress, update,
and recompress segments in the index. The downside is that deleted rows still take up
space in the index and are not immediately removed. Over time, the space consumed by
deleted rows may become nontrivial, at which point index maintenance (discussed in
Chapter 14) can periodically be used to reclaim this space.
Figure 4-6 adds the delete bitmap into the architecture diagram of columnstore indexes.

Figure 4-6. Adding the delete bitmap into the columnstore index architecture

Both the delete bitmap and the delta store are components of a columnstore index
that only exist when needed. If no deleted rows exist, then there will be no delete bitmap
to overlay onto the compressed rowgroups. Similarly, an empty delta store will not
need to be checked when queries are executed against the columnstore index. These
components exist when necessary, but otherwise will have no impact on performance
when not needed.

Nonclustered Columnstore Index Architecture

The architecture of nonclustered columnstore indexes is quite similar to that of clustered
indexes. Instead of being the primary storage mechanism for a table, the nonclustered
columnstore index is an additional index structure that exists alongside a clustered
rowstore index.
An important difference is that a column list may be specified on the nonclustered
columnstore index, allowing for more targeted optimization projects. Additionally, filters
may be applied to nonclustered columnstore indexes, allowing for data to be targeted
based on whether it is hot, warm, or cold, thus decreasing index size and the cost to write
to the index.

42
Chapter 4 Columnstore Index Architecture

Note the following brief definitions for the usage of hot/warm/cold data:

• Hot Data: Real time and actively used by a system with regular reads,
writes, and high concurrency. Availability is expected to be high and
latency should be very low.

• Warm Data: Represents data that is used less frequently. Availability

is still expected to be high, but there is a tolerance for more latency
when retrieving it.

• Cold Data: Represents infrequently accessed data that is old,

archived, or maintained for posterity. Availability is more flexible for
cold data and high latency is tolerable. Concurrency is very low for
cold data.

Organizations will label data in a range from hot to cold, and the specifics will
vary depending on how their data is used. Figure 4-7 shows the interaction between a
clustered rowstore index and a filtered nonclustered columnstore index.

Figure 4-7. A clustered rowstore index with a filtered nonclustered

columnstore index

Note that the purpose of a nonclustered columnstore index is to provide real-time

analytics against a transactional table, which requires careful consideration prior
to implementing. Chapter 12 will dive into more detail with regard to nonclustered
columnstore indexes including options, demonstrations, and optimal use cases.

Physical Data on Pages

Thus far, the discussion of architecture has occurred at a logical level. That is, rowgroups,
segments, delta store, and the delete bitmap have been presented as generic containers
for columns of data. The physical storage of columnstore indexes on pages helps in fully
appreciating how they can efficiently service analytic queries.
43
Chapter 4 Columnstore Index Architecture

All data in SQL Server that is not stored in memory-optimized structures is stored on
8 kilobyte pages. These pages reside on physical storage and are read into memory when
needed as-is. If a page is compressed, it remains compressed until its data is needed.
Whether a page is in a rowstore or columnstore index, it is read into memory in the same
fashion. The delete bitmap and delta store are also maintained on pages and read into
memory as needed.

Figure 4-8. Structure of a page in SQL Server

Figure 4-8 illustrates the basic structure of a page in SQL Server. The primary
components of pages are the

• Page header

• Data

• Row offsets
The page header contains basic information about the page such as the object that
owns it, the type of data stored in it, and the amount of unallocated space available for
additional data to be written.
The data rows are the actual data stored on the page. This may be the physical data
in a table (the clustered index or heap), index entries, or a variety of other contents that
are available under a variety of circumstances. For this discussion of columnstore and
rowstore indexes, the data and index data are all that is of immediate concern.
Row offsets store the position that each row starts, allowing SQL Server to locate the
data for any given row once a page is read into memory.
When data is stored in clustered indexes on a rowstore table, data is written to pages
row by row. That is, each column is written sequentially on the page for each row one
after another, as seen in Figure 4-9.

44
Chapter 4 Columnstore Index Architecture

Figure 4-9. Storage of rowstore data on a page in SQL Server

This table has six columns, and they are written sequentially for each row, one after
another. SQL Server will continue to write rows sequentially for this table to the same
page until it runs out of space, at which point a new page will be allocated to the table
and the data will continue there.
This structure is optimized for reading or writing small range lookups that consist
of a set of sequential rows. For example, a query that returned some (or all) of columns
from rows two through five would only need to read the single page shown in Figure 4-9.
If additional rows were required that were on other pages, then those pages would also
be read into memory to satisfy the query.
A clustered or nonclustered rowstore index writes its binary tree structure to pages
as a sequence of pointers. The root level is written first, followed by intermediate levels
that ultimately point to the underlying data at the leaf level of the index. Each level
of the index includes the clustered index keys, which are used to organize and link
between levels of the index. This structure is also optimized for seeking single or small
contiguous ranges of rows. For queries like this, less intermediate levels of an index need
to be read, reducing the number of pages needed to satisfy a query. Figure 4-10 shows a
visualization of a clustered rowstore (binary tree) index.

45
Chapter 4 Columnstore Index Architecture

Figure 4-10. Structure of a clustered rowstore index

Because this is the clustered index, the leaf levels contain the column data for each
row that is referenced by the clustered index. In a nonclustered index, the lowest level of
the index would contain pointers to the target data, to be used for key lookups. Similarly,
the leaf level of a nonclustered index would also contain any included columns that are
defined on the index.
A query that needed to read clustered index ID values between 25 and 75 could do
so by reading the index root page first and then pages 1 and 2. A query reading only
clustered index ID value 136 would read the index root and page 4.
In a columnstore index, there is no binary tree structure. A clustered columnstore
index is written as a sequence of compressed segments on as many pages as is needed
to support them. Figure 4-11 illustrates how this would look for a single page of a
columnstore index.

46
Chapter 4 Columnstore Index Architecture

Figure 4-11. Storage of columnstore data on a page in SQL Server

Note that this data, a subset of a compressed segment, contains sequential values for
the same column. This sequence would continue until the end of the rowgroup, at which
point a new segment would begin, or if it were the last of the rowgroups, the next column
would begin.
A query that aggregated values for the first 40 rows of column 1 could retrieve data
very efficiently and do so by simply reading this single page. In a rowstore index for
the same table, it would be necessary to scan data for all columns in those 40 rows and
read those pages into memory. For a table with 20 columns, that would require 20 times
the effort.

Summarizing Differences
The key to understanding the difference between the architecture of rowstore indexes
and columnstore indexes is to consider the logical storage of data vs. the physical storage
of data.
Rowstore indexes store data physically in order by row and then column. Data is
logically accessible as it is ordered by row. Therefore, this architecture is optimized for
queries that seek a sequence or limited subset of rows. Binary tree indexes provide a
speedy way to search for that data based on other columns that the clustered index does
not order the data by.

47
Chapter 4 Columnstore Index Architecture

Columnstore indexes store data physically grouped by column, but ordered by row.
This convention allows for filters to limit the number of rows returned and column
groupings to reduce the pages read as the number of columns needed for a query
decreases. This allows queries to read significantly more rows, but maintain excellent
performance when the column list is limited. Chapter 10 explores how the order of data
within a columnstore index can be controlled to add an additional dimension of filtering
so that data can be sliced both horizontally and vertically.

48
CHAPTER 5

Columnstore
Compression
There are many features that allow columnstore indexes to perform exceptionally well
for analytic workloads. Of those features, compression is the most significant driver
in both performance and resource consumption. Understanding how SQL Server
implements compression in columnstore indexes and how different algorithms are used
to shrink the size of this data allows for optimal architecture and implementation of
analytic data storage in SQL Server.

Basics of Columnstore Compression

Data written to segments within a columnstore index is heavily compressed prior to
storage. The delta store helps buffer incoming writes to a columnstore index, ensuring
that the frequency that segments need to be written is not high enough to negatively
impact server performance.
Without any additional optimizations, data stored by columns rather than rows
will be quite efficient to compress. This is because the values within a given column
will usually repeat often and therefore can be more easily compressed that data that
repeats less often. Typical analytic tables will contain many rows, but generally repetitive
columns, such as

• Entity types

• Money
• Dates

• Counts

• Repetitive dimensions
49
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_5
Chapter 5 Columnstore Compression

Consider the three columns illustrated in Figure 5-1.

Figure 5-1. Example of repeated values within a representative data set

Eleven rows are represented, of which each column contains repeated values. The
Date column contains only three distinct values, whereas Status contains four and Hours
contains six. This repetition tends to scale well. If this were the data from a rowgroup
containing one million rows, it would be reasonable to expect a column like Status to
only contain a small number of values that are repeated frequently throughout the data
set. The more repetitive the data is, the more effectively it will compress.
As is expected, more values are repeated within the values present in a single column
than across different columns. This is true for two primary reasons:

• Different data types will have vastly different values. For example,
the contents of a string column are not easily comparable to that of a
money column.

• Columns representing different dimensions will have different ranges

of values that correspond to their typical usage. For example, a
decimal column containing the tax rates charged on new cars would
contain a very different set of common values than a decimal column
containing the cost of those cars.

Columnstore indexes align perfectly with these conventions, whereas rowstore indexes
struggle because row and page compression are forced to take into account all columns in a
row or page, which will comprise a wide variety of data types and dimensions.
Note that the delta store is not columnstore compressed. This data structure needs to be
written to as quickly as possible at runtime, and therefore forgoing expensive compression
removes the time and overhead required to process writes against it. This will mean that
the delta store takes up more space than it may have if columnstore compression were
used. Since the delta store tends to be small in comparison to the compressed columnstore
index contents and contain a predictable count of rows, the effort needed to maintain it is
constant and does not grow over time. Depending on SQL Server version, delta rowgroups
may use row or page compression to limit their storage and memory footprint.

50
Chapter 5 Columnstore Compression

Columnstore Compression Algorithms

Columnstore indexes do not repeatedly store strings, decimals, or dates as is in
segments. Instead, SQL Server makes use of a dictionary encoding and value encoding
to shrink and compress data, even if the raw values are large.
Compression may be summarized in the following sequential steps:

1. Encode all values prior to compression.

2. Calculate optimal row order (aka Vertipaq optimization).

3. Individually compress each segment.

Encoding is an important step in the compression process. The following are details
that explain different encoding algorithms, how they are used, and how they impact
overall compression effectiveness.

V
alue Encoding
This process seeks to reduce the footprint of numeric columns by adjusting their storage
to be able to use smaller sized objects. For integer data, which consists of TINYINT,
SMALLINT, INT, and BIGINT, values are compressed by a variety of mathematical
transformations that can divide all values by a common divisor or subtract all values
with a common subtrahend. Consider a column that is a 4-bit integer data type and
contains the values shown in Table 5-1.

Table 5-1. Sample Integer Column Data to Be Compressed

Values Minimum Size in Bits

60 6
150 8
90 7
9000 13
630 10
300000 19

51
Chapter 5 Columnstore Compression

Columnstore compression seeks to find common divisors and immediately notes

that each number is divisible by 10. For this example, the base is 10 and each value will
be referenced using the maximum common negative exponent of that base that it is
divisible by, as seen in Table 5-2.

Table 5-2. Integer Data Compressed Using a Negative Exponent

Values Value *10-1 Minimum Size in Bits

60 6 3
150 15 4
90 9 4
9000 900 10
630 63 6
300000 30000 15

The next compression step is to take the smallest value in the list as the base, zero it
out, and reference other values by subtracting that base from them, as seen in
Table 5-3.

Table 5-3. Integer Data Compressed by Rebasing

Values Value 10-1 Value 10-1 - 6 Minimum Size in Bits

60 6 0 0
150 15 9 4
90 9 3 2
9000 900 894 10
630 63 57 6
300000 30000 29994 15

Reviewing the minimum space required to store the values, the original size was
63 bits. After the first transformation, the size is reduced to 42 bits, and after the last
transformation, it is reduced to 37 bits. Overall, the reduction in size for this set of values

52
Chapter 5 Columnstore Compression

was 38%. The level of value compression seen will vary based on the set of values and
can be significantly higher or lower. For example, a set of values that are all the same will
compress into a single zeroed-out reference, as seen in Table 5-4.

Table 5-4. Identical Integer Data Compressed by Rebasing

Values Value - 1700000 Original Size in Bits Reduced Size in Bits

1700000 0 22 0
1700000 0 22 0
1700000 0 22 0
1700000 0 22 0
1700000 0 22 0

While this example may seem extreme, it is not that unusual as dimensions often
contain repeated values that fall into limited ranges.
Note that real-world transformations of data that occur in columnstore indexes
are more complex than this, but the process to determine their details is essentially
the same.
Decimals are converted into integers prior to compression, at which point they
follow the same value encoding rules as integers would. For example, Table 5-5 shows
the compression process for a set of tax rates that are stored in decimal format.

Table 5-5. Value Compression for Decimals

Values Value * 100 Value / 5 Value / 5 - 70 Original Size in Bits Reduced Size in Bits

15.2 1520 304 234 40 8

8.8 880 176 106 40 7
4.3 430 86 16 40 5
3.5 350 70 0 40 0
6.25 625 125 55 40 6

53
Chapter 5 Columnstore Compression

The original data is stored as DECIMAL(4,2), which consumes 5 bytes (40 bits) per
value. By converting to integers prior to compression, multiple transformations are
applied that reduce storage from 200 bits to 26 bits. Decimals may not seem the most
likely candidates for effective compression, but once they are converted into integer
values, they can take advantage of any of the transformations that can be applied to
integers.

D
ictionary Encoding
String data is stored more efficiently using dictionary encoding. In this algorithm,
each distinct string value in a segment is inserted into a dictionary and indexed using
an integer pointer to the dictionary. This provides significant space savings for string
columns that contain repeated values. Consider a segment that contains foods of the
data type VARCHAR(50), as seen in Table 5-6.

Table 5-6. Sample Data Set Containing VARCHAR(50) String Data

Values Original Size in Bytes

Taco 4
Hamburger 9
Fish Fry 8
Taco 4
Hamburger 9
Taco 4
Triple-Layer Chocolate Cake 27
Triple-Layer Chocolate Cake 27
Triple-Layer Chocolate Cake 27
Cup of Coffee 13

The first step to compressing this data is to create an indexed dictionary, as seen in
Table 5-7.

54
Chapter 5 Columnstore Compression

Table 5-7. Example of a Dictionary That Is Generated

from VARCHAR Data

Index ID VARCHAR Value

0 Fish Fry
1 Hamburger
2 Taco
3 Triple-Layer Chocolate Cake
4 Cup of Coffee

Since the segment contains five distinct values, the index ID for the dictionary will
consume 3 bits. The resulting mapping in Figure 5-2 shows the resulting dictionary
compression of this data.

Figure 5-2. Sample dictionary compression of a set of ten strings

The original data required 132 bytes in storage, whereas the encoded data required
only 30 bits (3 bits per value), plus the space required to store the lookup values in the
dictionary. The dictionary itself consumes space for each distinct value in the segment
data, but as a bonus, a dictionary may be shared by multiple rowgroups. Therefore, a
table with 100 rowgroups may reuse dictionaries for each column across all of those
rowgroups.
The key to dictionary compression is cardinality. The less distinct values present
in the column, the more efficiently the data will compress. A CHAR(100) column with
5 distinct values will compress significantly better than a CHAR(10) column with 1000
distinct values. The dictionary size can be roughly estimated as the product of distinct

55
Chapter 5 Columnstore Compression

values and their initial size. Lower cardinality means that there are less string values to
store in the dictionary and the index used to reference the dictionary will be smaller.
Since the dictionary index IDs are repeated throughout rows in the table, a smaller size
can greatly improve compression.
Dictionaries can be viewed in SQL Server using the view sys.column_store_
dictionaries. The query in Listing 5-1 returns information about the dictionaries used in
the fact.Sale_CCI table created in Chapter 4.

Listing 5-1. Query to Return Dictionary Metadata

SELECT
       partitions.partition_number,
       objects.name AS table_name,
       columns.name AS column_name,
       CASE
              WHEN column_store_dictionaries.dictionary_id = 0 THEN 'Global
Dictionary'
              ELSE 'Local Dictionary'
       END AS dictionary_scope,
       CASE WHEN column_store_dictionaries.type = 1 THEN 'Hash dictionary
containing int values'
               WHEN column_store_dictionaries.type = 2 THEN 'Not used' --
Included for completeness
               WHEN column_store_dictionaries.type = 3 THEN 'Hash
dictionary containing string values'
               WHEN column_store_dictionaries.type = 4 THEN 'Hash
dictionary containing float values'
       END AS dictionary_type,
       column_store_dictionaries.entry_count,
       column_store_dictionaries.on_disk_size
FROM sys.column_store_dictionaries
INNER JOIN sys.partitions
ON column_store_dictionaries.hobt_id = partitions.hobt_id
INNER JOIN sys.objects
ON objects.object_id = partitions.object_id
INNER JOIN sys.columns

56
Chapter 5 Columnstore Compression

ON columns.column_id = column_store_dictionaries.column_id
AND columns.object_id = objects.object_id
WHERE objects.name = 'Sale_CCI';

The results of this query are shown in Figure 5-3.

Figure 5-3. Dictionary metadata results for the table fact.Sale_CCI

This view provides the cardinality of the column in entry_count, as well as the size of
the dictionary and its type. There are a number of noteworthy takeaways from this data:

• Dictionaries may be global or local. A local dictionary only applies to

a single rowgroup (or on rare occasions, rowgroups compressed via
the same process as that one). A global dictionary is shared across
many or all segments for a given column.

• Dictionaries may be created for numeric columns if SQL

Server believes that will provide better compression than value
compression. As a result, numeric columns can appear in this
metadata, but some will not. Typically, numeric columns with a low
cardinality will tend to use dictionary compression, whereas numeric
columns that have a high cardinality will use value compression. In
the preceding example, Sale Key, the primary key on the underlying
data, is value compressed, whereas Bill To Customer Key (containing
only three unique values) is dictionary compressed.

• Additional metadata may be gleaned from sys.columns, if needed,

such as the data type of the underlying data, its length, precision,
and scale.

57
Chapter 5 Columnstore Compression

This metadata provides detail about each dictionary in the columnstore index,
but does not directly relate segments to dictionaries. The query in Listing 5-2 links
sys.column_store_segments to sys.column_store_dictionaries to provide this added
information.

Listing 5-2. Query to Return More Detailed Dictionary Metadata

SELECT
       column_store_segments.segment_id,
       types.name AS data_type,
       types.max_length,
       types.precision,
       types.scale,
       CASE
              WHEN PRIMARY_DICTIONARY.dictionary_id IS NOT NULL THEN 1
              ELSE 0
       END AS does_global_dictionary_exist,
       PRIMARY_DICTIONARY.entry_count AS global_dictionary_entry_count,
       PRIMARY_DICTIONARY.on_disk_size AS global_dictionary_on_disk_size,
       CASE
              WHEN SECONDARY_DICTIONARY.dictionary_id IS NOT NULL THEN 1
              ELSE 0
       END AS does_local_dictionary_exist,
       SECONDARY_DICTIONARY.entry_count AS local_dictionary_entry_count,
       SECONDARY_DICTIONARY.on_disk_size AS local_dictionary_on_disk_size
FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id
INNER JOIN sys.objects
ON objects.object_id = partitions.object_id
INNER JOIN sys.columns
ON columns.object_id = objects.object_id
AND column_store_segments.column_id = columns.column_id
INNER JOIN sys.types
ON types.user_type_id = columns.user_type_id
LEFT JOIN sys.column_store_dictionaries PRIMARY_DICTIONARY

58
Chapter 5 Columnstore Compression

ON column_store_segments.primary_dictionary_id =
PRIMARY_DICTIONARY.dictionary_id
AND column_store_segments.primary_dictionary_id <> -1
AND PRIMARY_DICTIONARY.column_id = columns.column_id
AND PRIMARY_DICTIONARY.hobt_id = partitions.hobt_id
LEFT JOIN sys.column_store_dictionaries SECONDARY_DICTIONARY
ON column_store_segments.secondary_dictionary_id =
SECONDARY_DICTIONARY.dictionary_id
AND column_store_segments.secondary_dictionary_id <> -1
AND SECONDARY_DICTIONARY.column_id = columns.column_id
AND SECONDARY_DICTIONARY.hobt_id = partitions.hobt_id
WHERE objects.name = 'Sale_CCI'
AND columns.name = 'Bill To Customer Key';

Note that if a table is partitioned or has many columns, the result set could become
quite large, so this query is filtered by both table and column to focus on a single set of
segments, for demonstration purposes. The results help to fill in the blanks for dictionary
metadata, as seen in Figure 5-4.

Figure 5-4. Dictionary metadata detail for Sale_CCI. [Bill To Customer Key]

This detail differentiates between local and global dictionaries, as well as adds in
some column-level detail. Based on the results, the integer column [Bill To Customer
Key] compresses quite well as it contains only three distinct values in its dictionary.
Sixty-eight bytes for a dictionary is tiny and will provide immense savings for a column
that was previously 4 bytes per row across 25 million rows. Each value is reduced from 4
bytes to 2 bits, allowing it to be compressed at a ratio of 16:1!
It is important to note that a dictionary is limited in size to 16 megabytes. Since a
segment may only have one global and one local dictionary assigned to it, if a dictionary
reaches the 16MB threshold, then a rowgroup will be split into smaller rowgroups to
allow for each to possess dictionaries that are below this limit. If inserting new rows into

59
Chapter 5 Columnstore Compression

a table would cause any of its dictionaries to exceed the dictionary size cap, then a new
rowgroup will be created to accommodate new rows. There are two key takeaways from
this information:

1. Very long columns may force dictionary splits and result in

undersized rowgroups. This will reduce compression efficiency
and increase the number of dictionaries required to service
the table.

2. Columns with very high cardinality may also approach this

limit and result in dictionary splits that lead to the creation of
undersized rowgroups.

Numeric columns with high cardinality will typically use value compression to avoid
this problem, whereas string columns that are long and/or have a high cardinality may
not be able to avoid this issue. Sixteen megabytes may sound small when compared to
massive OLAP tables, but dictionaries themselves are compressed and typical analytic
data will not approach the dictionary size limit. Testing should always be used to confirm
either of these scenarios, rather than assuming they will or will not happen.

Should String Data Be Normalized?

Before advancing to a discussion of the remainder of the compression process, a
question that arises from the efficiency of dictionary compression is whether or not
normalization of dimensions is necessary or useful for columnstore compressed data
that uses dictionary compression.
In the world of transactional data, dimensions are often normalized. This saves space
and memory and allows for relational integrity, uniqueness, and other constraints to be
implemented at runtime. In analytic data, though, data validation is often implemented
via validation, rather than the use of triggers, keys, or check constraints. As a result,
normalization was often used more to save resources or provide convenient lookup
tables for use by visualization or analytics applications.
This question can be tested by creating a new version of Fact.Sale_CCI where
the Description column is normalized into a lookup table and referenced by a
SMALLINT key. Listing 5-3 provides definitions for this new table and its corresponding
lookup table.

60
Chapter 5 Columnstore Compression

Listing 5-3. Table Structure for Fact.Sale_CCI with a Normalized Description

CREATE TABLE Dimension.Sale_Description

(      Description_Key SMALLINT NOT NULL IDENTITY(1,1) PRIMARY KEY
CLUSTERED, [Description] NVARCHAR(100) NOT NULL);
CREATE TABLE Fact.Sale_CCI_Normalized
(      [Sale Key] [bigint] NOT NULL,
       [City Key] [int] NOT NULL,
       [Customer Key] [int] NOT NULL,
       [Bill To Customer Key] [int] NOT NULL,
       [Stock Item Key] [int] NOT NULL,
       [Invoice Date Key] [date] NOT NULL,
       [Delivery Date Key] [date] NULL,
       [Salesperson Key] [int] NOT NULL,
       [WWI Invoice ID] [int] NOT NULL,
       [Description Key] SMALLINT NOT NULL,
       [Package] [nvarchar](50) NOT NULL,
       [Quantity] [int] NOT NULL,
       [Unit Price] [decimal](18, 2) NOT NULL,
       [Tax Rate] [decimal](18, 3) NOT NULL,
       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,
       [Total Dry Items] [int] NOT NULL,
       [Total Chiller Items] [int] NOT NULL,
       [Lineage Key] [int] NOT NULL);

When populated with data, the Dimension.Sale_Description table is loaded first with
any new distinct Description values, and then Fact.Sale_CCI_Normalized is loaded with
data that joins the dimension table to retrieve values for Description_Key. Figure 5-5
shows the side-by-side comparison of the storage of these tables.

61
Chapter 5 Columnstore Compression

Figure 5-5. Storage comparison between normalized and denormalized columns

The table on the left is the original table, whereas the table on the right has
the Description column normalized into a SMALLINT. In this specific example,
normalization allowed for a significant space savings, about 15:1 compression savings
from the original table!
Generally speaking, normalization will help most when the number of distinct values
is low and the length of the columns is high, as dictionary encoding for a SMALLINT
column will produce superior compression than dictionary encoding for a large
VARCHAR column.
The downside to normalization is that data load processes will take longer. Joining
additional dimension columns at runtime to load data requires time and resources. The
decision as to whether to normalize or not should be carefully considered and testing
conducted to determine if sufficient gains are realized to justify the change.
Ultimately, business logic and organizational needs will drive this decision at a high
level, whereas technical considerations will allow for further tuning as needed. There is
no right or wrong answer to this question; therefore, consider either approach valid until
hands-on testing proves otherwise.

Row Order (Vertipaq) Optimization

Because a columnstore index is organized by column, rather than by row, it is possible
for SQL Server to rearrange the order of rows within each rowgroup at the time each
segment is compressed. This optimization is used whenever possible and attempts to
arrange rows such that like values are adjacent to each other.

62
Chapter 5 Columnstore Compression

Columnstore compression is highly sensitive to repeated sequences of values;

therefore, this optimization can provide exceptional compression and performance
improvements. Vertipaq optimization is the algorithm that seeks to rearrange rows in
the order that results in the most effective compression of the underlying data. This
algorithm is used in other Microsoft storage technologies, such as SQL Server Analysis
Services, PowerPivot, and PowerBI.
Vertipaq optimization is an expensive process, and the more columns a columnstore
index has, the longer it takes to find the optimal row order. Therefore, SQL Server
will limit the search time to ensure that building the index doesn’t take an excessive
amount of time. As a result, Vertipaq optimization will be somewhat less effective on
wider tables. This fact should not affect database architecture decisions, but can be a
useful troubleshooting step if a columnstore index on a wide table begins to experience
suboptimal compression as more columns continue to be added.
A critically important key to this optimization is that some architectural decisions
can prevent it from being used. There are two scenarios in which Vertipaq optimization
will not be used:

1. When the tuple mover merges rows from the delta store into a
clustered columnstore index that has at least one nonclustered
rowstore index

2. For columnstore indexes that are built on memory-optimized

tables

This does not mean that nonclustered rowstore indexes should never be used on
clustered columnstore indexes. Nor should memory-optimized columnstore indexes
be avoided. Instead, the fact that Vertipaq compression will not operate on these
structures should be an additional input into architectural decision-making process. If a
nonclustered rowstore index is required to service a specific set of critical queries, then
it should be used. If its use is optional, then consider the negative impact it may have on
compression.
While the details of how the algorithm works for a given columnstore index are not
exposed through any convenient views, it is possible to check whether or not a rowgroup
has Vertipaq optimization with the query in Listing 5-4.

63
Chapter 5 Columnstore Compression

Listing 5-4. Query That Returns Whether or Not Vertipaq Optimization Is Used

SELECT
       objects.name,
       partitions.partition_number,
       dm_db_column_store_row_group_physical_stats.row_group_id,
       dm_db_column_store_row_group_physical_stats.has_vertipaq_
optimization
FROM sys.dm_db_column_store_row_group_physical_stats
INNER JOIN sys.objects
ON objects.object_id =
dm_db_column_store_row_group_physical_stats.object_id
INNER JOIN sys.partitions
ON partitions.object_id = objects.object_id
AND partitions.partition_number =
dm_db_column_store_row_group_physical_stats.partition_number
WHERE objects.name = 'Sale_CCI'
ORDER BY dm_db_column_store_row_group_physical_stats.row_group_id;

The column has_vertipaq_optimization in sys.dm_db_column_store_row_

group_physical_stats indicates if that rowgroup was compressed using the Vertipaq
optimization or not. In general, expect to see this column return a value of 1, unless the
columnstore index is either on a memory-optimized table or is used in conjunction
with nonclustered rowstore indexes. If has_vertipaq_optimization is unexpectedly zero,
further research is warranted to understand why that is the case and determine what
changes were made that prevented this optimization from being used. Figure 5-6 shows
a sample of the results of the previous query.

64
Chapter 5 Columnstore Compression

Figure 5-6. Results of query to validate if Vertipaq optimization was used

In this scenario, Vertipaq optimization is in use for all rowgroups in this

columnstore index.

Other Compression Algorithms

After segments have been encoded and rowgroups have been reordered, the final step
in columnstore compression is to apply any remaining compression algorithms to the
resulting data. At this point, it is possible that no further compression gains are possible
for a given segment. This will be most often true with values that do not repeat. The
details of how columnstore indexes compress data are provided here to assist in more
advanced optimization of compression and performance, when possible.
One form of compression used by columnstore indexes is run-length encoding,
which seeks to group together like values and a count, rather than listing repetitive
values in a segment. Vertipaq optimization is a prerequisite to this step, and without it,
run-length encoding will be significantly less effective.
Consider the example of foods provided earlier in Figure 5-2. The encoded data has
five different values that have been dictionary encoded to 3-bit numeric values. Vertipaq
optimization would take the resulting data set and order it, as shown in Figure 5-7.

65
Chapter 5 Columnstore Compression

Figure 5-7. Vertipaq optimization of a set of encoded values

Note that the encoded values are now reordered so that each value is grouped next to
the same value. Run-length encoding groups together each like value logically to reduce
the overall amount of storage required for the segment. Logically, the resulting structure
can be illustrated as shown in Figure 5-8.

Figure 5-8. Run-length encoding applied to a numeric data set

Each value has been provided with a count, shown in parenthesis after it. There is a
single instance of the value 0, followed by two instances of 1, three instances of 2, three
instances of 3, and one instance of 4. This algorithm thrives on data that is repetitive and
works exceptionally well in conjunction with dictionary encoding and Vertipaq compression.
Bit array compression is used when a segment contains a small number of distinct
values and the data cannot benefit significantly from Vertipaq compression and run-length
encoding. A bit array is an array in which each distinct value is assigned a column and each
66
Chapter 5 Columnstore Compression

row is mapped to that column using bits. Figure 5-9 illustrates how this compression would
look when used on an unencoded and unsorted data set.

Figure 5-9. Bit array compression applied to an unencoded data set

While this may seem like a complicated transformation, the resulting ones and zeros
are quite compact and provide a structure that can be quickly decompressed.
There are a handful of other more complex compression algorithms that may be
used to progressively shrink the size of each segment of data. The algorithms used can
vary by segment and may be used often, sometimes, or not at all. When this process
is complete, the final level of compression is applied by SQL Server, which utilized its
xVelocity compression algorithm.
Columnstore segments are stored as Large Objects (LOB) in SQL Server. The details
of xVelocity compression are not public, and therefore we cannot delve further into how
they work. While the transformations used to convert the structures discussed thus far
in this chapter into their final form are not fully known, we can infer their effectiveness
by turning on STATISTICS IO and viewing the reads against a columnstore index when a
query is issued. Consider the simple analytic query in Listing 5-5.

Listing 5-5. Simple Analytic Query to Demonstrate Columnstore IO

SELECT
SUM(Quantity)
FROM fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/1/2016'
AND [Invoice Date Key] < '2/1/2016'

67
Chapter 5 Columnstore Compression

The result of this query is shown in Figure 5-10.

Figure 5-10. Sample query output

Clicking the Messages tab as shown in the results provides additional information on
the reads and writes performed by the query, as seen in Figure 5-11.

Figure 5-11. Statistics IO output for a sample query

There is quite a bit of information returned, but for the moment, the LOB logical
reads circled in red provide an indication of the reads required to satisfy this query.
Columnstore indexes will not report reads as logical reads, but instead as LOB logical
reads, which indicates that their segments are stored as Large Objects and not as
traditional SQL Server data structures. If this number of reads seems high, that is
because it is! Later chapters in this book will provide additional optimizations that will
reduce storage and memory footprint, as well as greatly improve query performance.

Columnstore Archive Compression

SQL Server provides an additional level of compression that can be used to further
shrink the footprint of columnstore index data. While archive compression can provide
a higher compression ratio than standard columnstore compression, it costs more

68
Chapter 5 Columnstore Compression

computing resources to read and write the data. As a result, this option should be
reserved for data that is not frequently used, such as

• Cold or warm storage

• Data that is rarely accessed

• Data used for applications where performance is not critical

Columnstore archive compression uses the Microsoft Xpress compression algorithm.

While this algorithm is complex, it is publicly documented and is not a Microsoft trade
secret. This algorithm relies on compressing data with repeated byte sequences and
combines the LZ77 and Huffman compression algorithms to further shrink the size of
columnstore data. The details of these compression algorithms are too extensive for the
pages of this book, but are publicly available to anyone interested.
The type of compression used in a columnstore index may be specified on a
partition-by-partition basis, allowing older/less used data to effectively use archive
compression, whereas newer data can use standard columnstore compression.
Chapter 11 delves into how partitioning can be combined with columnstore indexes to
improve data storage, maintenance, and performance.
To demonstrate the effectiveness of columnstore archive compression, a new version
of the Sale_CCI table will be created, as seen in Listing 5-6.

Listing 5-6. Table Created with Columnstore Archive Compression

CREATE TABLE Fact.Sale_CCI_Archive

69
Chapter 5 Columnstore Compression

[Tax Rate] [decimal](18, 3) NOT NULL,

       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,
       [Total Dry Items] [int] NOT NULL,
       [Total Chiller Items] [int] NOT NULL,
       [Lineage Key] [int] NOT NULL);
CREATE CLUSTERED COLUMNSTORE INDEX CCI_Sale_CCI_Archive ON
fact.Sale_CCI_Archive WITH (DATA_COMPRESSION=COLUMNSTORE_ARCHIVE);

Once populated with data, this table can be compared to the previous version that
used standard columnstore compression, as seen in Figure 5-12.

Figure 5-12. Comparison of columnstore and columnstore archive compression

Columnstore archive compression provided an immense space savings when

compared to standard columnstore compression. This may seem absurdly impressive,
but some of this savings can be achieved by optimizing data order, which will be
discussed in detail in Chapter 10.
In practice with typical production data loads, archive compression will deliver an
additional 15–30% savings over standard columnstore compression. For a large analytic
table, this can equate to a significant reduction in its data footprint! In general, only
consider archive compression for data that will not be written to or read often, as the cost
to compress and decompress it is nontrivial. For data that is rarely accessed, it is an easy
way to save on both storage and memory resources.

70
Chapter 5 Columnstore Compression

The Compression Life Cycle

A key to all compression within SQL Server is that compressed pages remain compressed
until needed by a query. Therefore, pages are compressed on storage and remain
compressed there for the lifetime of an index. When pages are read into memory, they
remain compressed. Therefore, any storage space saved by the use of compression will
equate into memory savings when the data is needed.
This fact is critical to the efficiency of columnstore indexes. Compressed segments
remain compressed on disk as well as in memory and are not decompressed until a
query requires their data. Any optimizations that further decrease the size of compressed
segments will translate directly into memory savings. The remainder of this book will
focus on optimizations, best practices, and features that allow columnstore indexes to be
used as efficiently as possible, thus making the most of their compression and ensuring
they can scale for any data, regardless of how big it gets.

71
CHAPTER 6

Columnstore Metadata
Each compressed segment within a columnstore index not only stores analytic data,
but through metadata can describe its contents with more precision than rowstore
tables can.
This metadata resides in system views and allows SQL Server to make intelligent
query processing decisions at runtime. This chapter dives into this metadata in
detail, how SQL Server uses it, and how it can be used to improve the performance of
columnstore indexes.

Available Columnstore Metadata

Metadata is available through a handful of system views that provide overview-level data
about segments and rowgroups, as well as detail data that describes physical properties,
dictionaries, and operational metrics. The following is a brief review of each view and its
contents.

R
owgroup Metadata
Each compressed rowgroup contains up to 220 rows whose contents are typically static
outside of the influence of index maintenance. Because of this, the row counts, deleted
row counts, size, and other details are accurate and can be used to understand the size
and shape of the underlying data.
The query in Listing 6-1 returns the table and index name, as well as all columns
from the view sys.column_store_row_groups.

73
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_6
Chapter 6 Columnstore Metadata

Listing 6-1. Query to Return Metadata for All Rowgroups in a Columnstore Index

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       column_store_row_groups.*
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
WHERE tables.name = 'Sale_CCI'
ORDER BY tables.object_id, indexes.index_id,
column_store_row_groups.row_group_id;

Figure 6-1 contains a subset of the rowgroup metadata returned.

Figure 6-1. Rowgroup metadata for the columnstore index on Sale_CCI

Each row returned by sys.column_store_row_groups represents the set of rows

contained within a rowgroup, regardless of how many columns are in the table. A table
with ten columns will contain ten segments per rowgroup. Similarly, a table with 30
columns will contain 30 segments per rowgroup. The following is a brief overview of
each column in the view.

P
artition_number
A rowgroup cannot span multiple partitions; therefore, a partitioned table will have
separate sets of rowgroups in each partition, without overlap between them. This
column provides the partition that a given rowgroup resides on.

74
Chapter 6 Columnstore Metadata

Delta_store_hobt_id
For open rowgroups in the delta store, an ID is provided here that links to
sys.internal_partitions and represents the delta rowgroup data it contains. A standard
compressed rowgroup will contain NULL for this column.

State and State_Description

This describes the current status of the rowgroup, whether it is OPEN and accepting new
rows, CLOSED, but not yet compressed by the tuple mover, COMPRESSED, or INVISIBLE
(a rowgroup in the process of being created).

Total_rows
This provides the row count for the rowgroup. Summing this value across all rowgroups
allows for an accurate row count to be collected for the table without querying it directly.
If many rowgroups have low row counts, it is indicative of a process problem worthy of
investigation.

Deleted_rows
If any rows are flagged as deleted in the delete bitmap for this rowgroup, the count of
deleted rows is provided here.

Size_in_bytes
The total size of the rowgroup, which is the sum space used by all segments residing
in that rowgroup. This can be used to quickly determine the size of all or part of a
columnstore index.

Segment Metadata
The basic unit of storage in a columnstore index is the segment. Each segment
represents a single column and its contents for the set of rows contained in each
rowgroup. The number of segments that comprise a columnstore index is the product of
the number of rowgroups in the index and columns in the table.

75
Chapter 6 Columnstore Metadata

Sys.column_store_segments provides extensive detail on the contents of each

segment. This data is accurate and updated whenever data within the segment changes,
unlike SQL Server statistics which require periodic updating to ensure a reasonable
approximation of reality.
The data in this view is more than a reference for curious administrators. It is also
used by the query optimizer to make intelligent decisions about what data to read at
runtime when a columnstore index is queried. This allows segments to be skipped when
the metadata indicates they are not needed to satisfy a query. This can significantly
reduce reads and improve query performance in ways that are unavailable to rowstore
tables in SQL Server.
The query in Listing 6-2 returns a variety of useful columns from this view.

Listing 6-2. Query That Returns Columnstore Segment Metadata

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       columns.name AS column_name,
       partitions.partition_number,
       column_store_segments.encoding_type,
       column_store_segments.row_count,
       column_store_segments.has_nulls,
       column_store_segments.base_id,
       column_store_segments.magnitude,
       column_store_segments.min_data_id,
       column_store_segments.max_data_id,
       column_store_segments.null_value,
       column_store_segments.on_disk_size
FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id
INNER JOIN sys.indexes
ON indexes.index_id = partitions.index_id
AND indexes.object_id = partitions.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.columns
76
Chapter 6 Columnstore Metadata

ON tables.object_id = columns.object_id
AND column_store_segments.column_id = columns.column_id
WHERE tables.name = 'Sale_CCI'
ORDER BY columns.name, column_store_segments.segment_id;

The segment metadata returned can be seen in Figure 6-2.

Figure 6-2. Segment metadata for the columnstore index on Sale_CCI

Details about columnstore segments delve into how they are compressed and stored,
which can provide valuable clues to the efficiency of columnstore compression for
each column within a columnstore index. The following is detail about some of the key
columns in this view.

E ncoding_type
This indicates if this segment is encoded using a dictionary and the type of data stored in
it, as follows:

• 1: Nonstring/nonbinary data with no dictionary. The values in this

segment are transformed using base/magnitude encoding.
• 2: Nonstring/nonbinary data with a dictionary.

• 3: String or binary data with a dictionary.

• 4: Nonstring/nonbinary data with no dictionary. The values in this

segment are stored as is, with no additional transformations.

• 5: String or binary data with no dictionary.

Typically, segments with many distinct values (aka high cardinality) will tend to use
value-based encoding and not implement a dictionary lookup, whereas segments with
low cardinality will more often use a dictionary to reduce the storage of repeated values.
This decision is made internally by SQL Server to improve compression and reduce the
space consumed by a segment.
77
Chapter 6 Columnstore Metadata

R
ow_Count
This row count matches the row count of the corresponding rowgroup and represents
the number of values (repeated or distinct) contained within a columnstore segment.

H
as_nulls
If a segment contains at least one NULL, then this is set to 1; otherwise, it is 0. Note that
this column does not indicate if a column allows NULL, but instead reports on whether a
segment happens to have NULL within its set of values. Therefore, different segments for
the same column can have different values for this bit.

B
ase_id and Magnitude
These columns report directly on value-based encoding. If values were modified using a
base and exponent to reduce their storage size, that detail is represented here. The base
and magnitude can vary segment to segment and do not need to be the same for each
segment for any one column.

M
in_data_id and Max_data_id
This incredibly useful pair of columns provides the minimum and maximum values for a
segment, or the dictionary lookup values, if a dictionary is used. If transformations have
been made for value-based encoding, then the values provided by these columns will
include those modifications.
The minimum and maximum values for a segment are used by the query optimizer
to skip unnecessary rowgroups when reading a columnstore index. For example, if a
segment contains a minimum value of 1 and a maximum value of 400, then queries
that filter for values outside of this range can skip this rowgroup altogether. This
optimization is known as segment elimination and is key to the performance of large
columnstore indexes. Chapter 10 discusses segment elimination in detail and provides
the conventions needed to take full advantage of this feature.

N
ull_value
If a segment contains NULLs, then the value provided by this column is the numeric
representation for NULL within the encoding of the column.

78
Chapter 6 Columnstore Metadata

O
n_disk_size
The space consumed by the columnstore segment is provided by the column
on_disk_size, allowing for the space consumed by a columnstore index to be easily
broken down into granular details. If one segment out of many has an unusually large
size, there can be value in exploring why and determining if further optimizations can be
made to improve data compression.
This granular detail also allows for the space consumed by each column to be
calculated, as well as space consumed per column per partition. If a columnstore index
is growing large unusually quickly, this metadata allows for the source of growth to be
isolated to a column and set of segments to determine the source of that growth.

Rowgroup Physical Metadata

SQL Server provides the view sys.dm_db_column_store_row_group_physical_stats as
a reference to each rowgroup. Included are details regarding their current state and
how rowgroups were created. This information is invaluable when troubleshooting
unexpected performance challenges or when trying to understand why a columnstore
index rowgroups were built in a particular way.
The query in Listing 6-3 returns key information from this view, as well as some
associated metadata.

Listing 6-3. Query That Returns Columnstore Rowgroup Physical Details

SELECT
       objects.name AS table_name,
       indexes.name AS index_name,
       dm_db_column_store_row_group_physical_stats.partition_number,
       dm_db_column_store_row_group_physical_stats.row_group_id,
       dm_db_column_store_row_group_physical_stats.state_desc,
       dm_db_column_store_row_group_physical_stats.total_rows,
       dm_db_column_store_row_group_physical_stats.deleted_rows,
       dm_db_column_store_row_group_physical_stats.size_in_bytes,
       dm_db_column_store_row_group_physical_stats.trim_reason_desc,
       dm_db_column_store_row_group_physical_stats.transition_to_
compressed_state_desc,

79
Chapter 6 Columnstore Metadata

dm_db_column_store_row_group_physical_stats.has_vertipaq_
optimization,
dm_db_column_store_row_group_physical_stats.created_time
FROM sys.dm_db_column_store_row_group_physical_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_physical_stats
.object_id
INNER JOIN sys.indexes
ON indexes.object_id = dm_db_column_store_row_group_physical_stats
.object_id
AND indexes.index_id = dm_db_column_store_row_group_physical_stats.index_id
WHERE objects.name = 'Sale_CCI';

The results provide additional information about rowgroups, as seen in Figure 6-3.

Figure 6-3. Rowgroup physical stats for the columnstore index on Sale_CCI

This information provides clues as to how and when rowgroups were built. The
following are descriptions of the key columns presented in Figure 6-3.

S
tate_desc
This is the current state of the rowgroup, which is identical to the value found in
sys.column_store_row_groups. A rowgroup with a state of TOMBSTONE indicates
that it previously was part of a delta store and has been transitioned into compressed
rowgroups, but has yet to be cleaned up. This cleanup occurs asynchronously and
automatically, without the need for operator intervention.

Total_rows, Deleted_rows, Size_in_bytes

These are the same as their corresponding values in sys.column_store_row_groups and
are provided in this view for convenience.

80
Chapter 6 Columnstore Metadata

Trim_reason_desc
A rowgroup may contain up to 220 rows. There are many reasons why a rowgroup is
created and compressed with less than its full complement of rows. If this column shows
“NO_TRIM” for a rowgroup, then that indicates that it is a full rowgroup with 1,048,576
rows. If the rowgroup contains less than its maximum possible size in rows, then the trim
reason will explain why.
Typically, a columnstore index should mostly contain rowgroups with row counts
near 1,048,576. In the event that this is not the case, and a majority of rowgroups are
undersized, then understanding why is key to improving columnstore storage and
performance. There is no firm threshold for defining what “undersized” is, but typically
an undersized rowgroup contains less than 900,000 rows, or less than about 10% of its
maximum size.
Common reasons for a rowgroup to be trimmed include

• BULKLOAD: A bulk load process limited the resulting row count

to less than its maximum size. This is normal and rarely a cause for
concern.

• REORG: An index reorganize process was used to force compression

of a rowgroup. Since this is a manual process, it is easy to control,
if needed.

• DICTIONARY_SIZE: A dictionary is limited to 16 megabytes of space.

If a dictionary exceeds that size, then the rowgroup it corresponds
to will be split into another rowgroup. If this happens often and
if the resulting row counts are low, it indicates that dictionary
compression is harming compression performance. Normalizing a
large dimension into a lookup table is an effective way to resolve this
problem.

• MEMORY_LIMITATION: This indicates that SQL Server ran low on

memory while compressing the rowgroup and could not include all
rows. This is indicative of memory pressure, and research should
be conducted to determine why memory was low and if additional
memory should be added to the SQL Server to resolve this problem.
It is likely that memory pressure of this nature would affect other
processes on the server as well and alleviating it would be beneficial
to them, in addition to the columnstore index build process.
81
Chapter 6 Columnstore Metadata

• RESIDUAL_ROW_GROUP: When an index build operation

completes, leftover rows will be compressed into rowgroups using
this trim reason. This is normal and does not require investigation.

• AUTO_MERGE: This results when the tuple mover runs and merges
multiple rowgroups together. This is a good thing as it contributes to
larger rowgroups and the removal of undersized ones.

Transition_to_compressed_state_desc
This column describes how a rowgroup was moved from the delta store into a set of
compressed segments within a columnstore index. It provides additional information
for how the rowgroup was created and can be used in conjunction with the trim reason
to understand how all rowgroups were created, whether trimmed or not. The possible
values for the transition to compressed state description include

• NOT_APPLICABLE: This is found either in delta store rowgroups

or for rowgroups compressed prior to SQL Server 2016, when this
history was not maintained.

• INDEX_BUILD: This rowgroup was compressed via an index build

process.

• TUPLE_MOVER: This rowgroup was compressed by the tuple mover

after it inserted data into it from the delta store.

• REORG_NORMAL: An index reorganize process was executed that

moved a closed rowgroup from the delta store into the columnstore
index. This occurred after it was closed with its full number of rows
but before it was moved asynchronously via the tuple mover.

• REORG_FORCED: An index reorganize process was executed that

moved an open rowgroup from the delta store into the columnstore
index. This occurred before it had its full number of rows and would
normally be moved asynchronously via the tuple mover.

• BULKLOAD: A minimally logged bulk load process created the

rowgroup without using the delta store.

• MERGE: The tuple mover merged multiple rowgroups together into

this compressed rowgroup.

82
Chapter 6 Columnstore Metadata

Note that with the exception of the two transition reasons related to index reorganize
operations, the remaining reasons indicate the creation of compressed rowgroups
through automatic processes in SQL Server. These reasons should be seen as part
of normal operations and only investigated if an unsolved performance problem is
encountered.

Has_vertipaq_optimization
This indicates that the rows within a rowgroup could be reordered via Vertipaq
optimization to improve the compression ratio for this rowgroup.
If zero is indicated by this column, it is worth investigation. The space savings
associated with Vertipaq optimization is nontrivial and will result in faster queries
and less memory and storage consumption by a columnstore index. As discussed in
the previous chapter, the two reasons why Vertipaq optimization may not occur are as
follows:

1. The columnstore index resides on a memory-optimized table.

2. The columnstore index contains nonclustered rowstore indexes.

Note that if a nonclustered rowstore index is removed from a clustered columnstore

index, the Vertipaq optimization status will not change until an index rebuild is
executed.
Consider the benefits and drawbacks of architectural decisions that lead to Vertipaq
optimization not being used, as there are valid reasons to have a columnstore index in
memory, or to include nonclustered rowstore indexes.
One solution for a columnstore index that requires a supporting nonclustered
rowstore index is to implement a nonclustered index on the partitions with the data
relevant to the queries requiring the index. Vertipaq optimization will only be impacted
on rowgroups on partitions that contain the nonclustered index. Typically, data can be
separated into hot, warm, and cold data, depending on common workloads. Secondary
indexes can be split across partitions to target those different workloads effectively.
Similarly, filtered nonclustered rowstore indexes may be used to further limit the size
of the index and reduce its scope and impact on compression and storage.

83
Chapter 6 Columnstore Metadata

C
reated_time
This indicates the time that a rowgroup was created. This can help in understanding
how many rowgroups are created over time and the number of rows being added to the
columnstore index per unit time. In addition, created_time can be correlated to
size_in_bytes, allowing the increase in size of the columnstore index to be measured over
time in bytes.

Rowgroup Operational Statistics

SQL Server maintains cumulative details on the usage of columnstore indexes and
provides those details at the rowgroup level. This allows for an understanding of which
rowgroups within a columnstore index constitute hot, warm, or cold data and for
granular detail as to how each rowgroup is used.
The query in Listing 6-4 returns operational data for a single columnstore index.

Listing 6-4. Query That Returns Columnstore Rowgroup Operational Statistics

SELECT
       objects.name AS table_name,
       indexes.name AS index_name,
       dm_db_column_store_row_group_operational_stats.row_group_id,
       dm_db_column_store_row_group_operational_stats.index_scan_count,
       dm_db_column_store_row_group_operational_stats.scan_count,
       dm_db_column_store_row_group_operational_stats.delete_buffer_
scan_count,
       dm_db_column_store_row_group_operational_stats.row_group_lock_count,
       dm_db_column_store_row_group_operational_stats.row_group_lock_
wait_count,
       dm_db_column_store_row_group_operational_stats.row_group_lock_
wait_in_ms,
       dm_db_column_store_row_group_operational_stats.returned_row_count
FROM sys.dm_db_column_store_row_group_operational_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_operational_stats
.object_id

84
Chapter 6 Columnstore Metadata

INNER JOIN sys.indexes

ON indexes.object_id = dm_db_column_store_row_group_operational_stats
.object_id
AND indexes.index_id = dm_db_column_store_row_group_operational_stats
.index_id
WHERE objects.name = 'Sale_CCI';

The results in Figure 6-4 show individualized operational data for each rowgroup.

Figure 6-4. Rowgroup operational stats for the columnstore index on Sale_CCI

All metrics in this view are cumulative since the last restart of the SQL Server service.
Therefore, to meaningfully use this data, it must be captured periodically to calculate the
difference in measurements from one sample to the next.
The following is a subset of columns that can be used to track detailed columnstore
rowgroup usage.

I ndex_scan_count
This counts the number of times the columnstore index was scanned, regardless of
which rowgroups were requested. This value will be identical for all rowgroups in a given
partition.

S
can_count
This only counts the number of times this rowgroup was scanned. Comparing this to
index_scan_count allows for an understanding of how often a particular rowgroup is
needed to satisfy queries against the columnstore index.
In typical OLAP data, older data will be required less often than newer data, and
therefore rowgroups containing older data will see less scans than newer rowgroups.
The scan count columns in this view can provide the necessary data to support storage

85
Chapter 6 Columnstore Metadata

changes for cold or warm data. If older data is needed for posterity but rarely queried,
it can likely be off-loaded to slower and cheaper storage. Similarly, if there is a portion of
data that is newer and critically important, it may benefit from faster storage.

Delete_buffer_scan_count
This counts the number of times that the delete bitmap was consulted in order to
complete a query against this rowgroup. The delete buffer refers to both the delete
bitmap as it is stored alongside the columnstore index as a b-tree, as well as the in-
memory hash-table representation of that index.
Typically, deleted rows in a columnstore index are not problematic until their
quantity become large when compared to the size of the index. If delete_buffer_scan_
count is a larger number that is close in value to scan_count, then an index rebuild
may be worthwhile to remove the deleted rows from the index. Check the value for
deleted_rows in dm_db_column_store_row_group_physical_stats first to validate that
the number of deleted rows is indeed large before proceeding with index maintenance.
While an index rebuild can be executed as an online operation, it is still computationally
expensive and can be time-consuming for a large columnstore index.

Row_group_lock_count, Row_group_lock_wait_count,
and Row_group_lock_wait_in_ms
These provide the count (and details) of lock requests against this rowgroup. These are
typically the result of columnstore index writes colliding with reads. Rowgroup locks are
most frequent against rowgroups that are actively being written, whereas locks against
older rowgroups are uncommon.
Rowgroup locking is not inherently a bad thing, but if severe contention is occurring
between data loads and analytic processes, then understanding which rowgroups are
causing contention can assist in troubleshooting that situation.
Update operations will be the most likely to cause contention as they require writing
both to the delete bitmap and to the delta store. While updating a single row would
impact two rowgroups (one for the delete bitmap and one for the delta store), a larger
update might impact many rowgroups. In these scenarios, the locking caused by the
update can be quite disruptive.

86
Chapter 6 Columnstore Metadata

Similarly, if data is trickled into a columnstore index throughout the day, there is
more opportunity for contention than if it is managed via a single centralized process
less frequently and during a nonpeak time for analytics.
As is the case for many performance challenges, investigate contention when
needed. If row_group_lock_count appears high, but analytics speeds (and end-user
happiness) are good, then that fact is best saved for posterity, but not acted on yet.
Always cross-reference Row_group_lock_wait_count and Row_group_lock_wait_in_ms to
determine if locking resulted in waits or not. If the count of waits or the wait time is not
high, then further action is likely not needed. If unsure of whether or not the aggregate
wait time is high or not, consider the lock wait time per lock incident, as given by the
query in Listing 6-5.

Listing 6-5. Formula That Calculates the Lock Time per Lock Incidence

SELECT
       objects.name AS table_name,
       indexes.name AS index_name,
       dm_db_column_store_row_group_operational_stats.row_group_id,
       dm_db_column_store_row_group_operational_stats.scan_count,
       dm_db_column_store_row_group_operational_stats.row_group_lock_
wait_count,
       dm_db_column_store_row_group_operational_stats.row_group_lock_
wait_in_ms,
       CASE
              WHEN dm_db_column_store_row_group_operational_stats.row_
group_lock_wait_count = 0 THEN 0
              ELSE CAST(CAST(dm_db_column_store_row_group_operational_
stats.row_group_lock_wait_in_ms AS DECIMAL(16,2)) /
                      CAST(dm_db_column_store_row_group_operational_stats
.row_group_lock_wait_count AS DECIMAL(16,2))
AS DECIMAL(16,2))
       END AS lock_wait_ms_per_wait_incidence
FROM sys.dm_db_column_store_row_group_operational_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_operational_stats
.object_id

87
Chapter 6 Columnstore Metadata

INNER JOIN sys.indexes

ON indexes.object_id = dm_db_column_store_row_group_operational_stats
.object_id
AND indexes.index_id = dm_db_column_store_row_group_operational_stats
.index_id
WHERE objects.name = 'Sale_CCI';

The details returned provide an understanding of

• How many of the operations against a rowgroup resulted

in locking/waits?

• How long were waits due to locking when they occurred?

• How does wait time compare to overall rowgroup performance?

If contention is a problem within a columnstore index, consider streamlining data

load processes. Break update operations into sequences of deletes and inserts. This
reduces transaction size during data loads, speeds up update operations, and reduces
the opportunity for contention to be significant. Utilize bulk load processes via large
insert batches when possible as they will be faster, generate less contention, and reduce
transaction log size. Lastly, ensure that writes against the columnstore index are reduced
and centralized as much as possible. If intermediary write operations can be made
against a temporary table prior to modifying columnstore index data, then the result
will be less contention on the data used for analytics and faster data load speeds. Best
practices such as these are discussed in more detail in Chapter 15.

R
eturned_row_count
One additional metric provided that can assist in quantifying utilization is the overall
count of rows returned. This provides an additional dimension as to how much data is
read from this rowgroup over a period of time and can be compared to other rowgroups
in the index. Note that a lower row count may be the result of the rowgroup having less
rows and not simply less usage.

88
Chapter 6 Columnstore Metadata

Columnstore Index Memory Usage

The system view sys.dm_column_store_object_pool provides details about the memory
allocated to columnstore indexes. This can assist in troubleshooting memory pressure,
as well as capacity planning when creating a columnstore index table. This is a shared
view across an entire server; therefore, filtering on the database name is useful when
viewing metadata solely for a single database.
The query in Listing 6-6 returns details about memory usage for one specific
columnstore index.

Listing 6-6. Columnstore Memory Usage by the Sale_CCI Table

SELECT
       databases.name,
       objects.name,
       indexes.name,
       columns.name,
       dm_column_store_object_pool.row_group_id,
       dm_column_store_object_pool.object_type_desc,
       dm_column_store_object_pool.access_count,
       dm_column_store_object_pool.memory_used_in_bytes,
       dm_column_store_object_pool.object_load_time
FROM sys.dm_column_store_object_pool
INNER JOIN sys.objects
ON objects.object_id = dm_column_store_object_pool.object_id
INNER JOIN sys.indexes
ON indexes.object_id = dm_column_store_object_pool.object_id
AND indexes.index_id = dm_column_store_object_pool.index_id
INNER JOIN sys.databases
ON databases.database_id = dm_column_store_object_pool.database_id
LEFT JOIN sys.columns
ON columns.column_id = dm_column_store_object_pool.column_id
AND columns.object_id = dm_column_store_object_pool.object_id
WHERE objects.name = 'Sale_CCI'
AND databases.name = DB_NAME()
ORDER BY dm_column_store_object_pool.row_group_id, columns.name;

89
Chapter 6 Columnstore Metadata

The results are shown in Figure 6-5.

Figure 6-5. Sample of memory consumption by a columnstore index

Note the level of detail returned by sys.dm_column_store_object_pool. The results

indicate a query was recently run that targeted Invoice Date Key, as well as Quantity.
Because a columnstore index stores each column separately in their own segments, it
is only necessary to read segments into memory for the requested columns. This differs
from rowstore indexes, where pages contain all columns in the index and must be read
together.
In addition to column segments, memory is consumed by delete bitmaps and
dictionaries, allowing for a granular view of the memory consumed by each object
within a columnstore index. The following is a description of each new column returned
in Figure 6-5.

O
bject_type_desc
This is the type of columnstore object residing in memory. It will be one of the
following values:
• COLUMN_SEGMENT: Each row with this type represents a single
columnstore segment. These entries will contain a value for column_
id, allowing for further detail to be returned regarding the column
that the segment belongs to.

• PRIMARY_DICTIONARY, SECONDARY_DICTIONARY, and

BULKINSERT_DICTIONARY: These represent dictionaries
used for segments that are encoded with dictionary-based
encoding. Typically, the space consumed by dictionaries and their
corresponding column segments will be less than if the column were
stored using value-based encoding.

90
Chapter 6 Columnstore Metadata

• DELETE_BITMAP: This represents memory consumed by delete

bitmap objects. There will always be one delete bitmap per partition
that spans all rowgroups in a columnstore index on that partition.
The entries for each rowgroup indicate the granular details for the
part of the delete bitmap associated with that particular rowgroup.
These objects are present as containers, even if there are no rows
indicated as deleted within them.

A
ccess_count
This is a count of all read and write operations to this object in memory. This provides a
rough measure of how much use the object has gotten and which objects in memory are
used often vs. those that are used rarely.
Because this view examines objects residing in memory, the data is transient and
maintains history only for as long as an object is in memory. Therefore, access counts are
only cumulative since the time given by object_load_time.

O
bject_load_time
This is the time that this object was read into the object pool in memory. Combined with
access_count, the number of read or write operations per unit time can be calculated,
which may be useful in gauging how much usage particular objects are getting
in memory.
The data provided in dm_column_store_object_pool can also be aggregated so that
overall memory consumption by a given rowgroup, column, index, or object type can
be measured. For example, the query in Listing 6-7 quantifies how much memory is
consumed by each rowgroup.

Listing 6-7. Columnstore Memory Consumption by Rowgroup

SELECT
       databases.name,
       objects.name,
       indexes.name,
       dm_column_store_object_pool.row_group_id,
       SUM(dm_column_store_object_pool.access_count) AS access_count,
       SUM(dm_column_store_object_pool.memory_used_in_bytes) AS
memory_used_in_bytes
91
Chapter 6 Columnstore Metadata

FROM sys.dm_column_store_object_pool
INNER JOIN sys.objects
ON objects.object_id = dm_column_store_object_pool.object_id
INNER JOIN sys.indexes
ON indexes.object_id = dm_column_store_object_pool.object_id
AND indexes.index_id = dm_column_store_object_pool.index_id
INNER JOIN sys.databases
ON databases.database_id = dm_column_store_object_pool.database_id
LEFT JOIN sys.columns
ON columns.column_id = dm_column_store_object_pool.column_id
AND columns.object_id = dm_column_store_object_pool.object_id
WHERE objects.name = 'Sale_CCI'
AND databases.name = DB_NAME()
GROUP BY databases.name, objects.name, indexes.name,
dm_column_store_object_pool.row_group_id
ORDER BY dm_column_store_object_pool.row_group_id;

Figure 6-6 shows the results.

Figure 6-6. Memory consumption grouped by rowgroup

Depending on the index, its metadata, and its usage, some rowgroups may use
significantly more memory than others. In the example results in Figure 6-6, memory
is used relatively equally per rowgroup, indicating that queries are reading data evenly
from each rowgroup.

92
Chapter 6 Columnstore Metadata

Internal Columnstore Index Objects

There is one additional view that provides internal details about the structure of a
columnstore index, sys.internal_partitions. The query in Listing 6-8 returns details about
any internal partitions associated with the rowgroups in a columnstore index.

Listing 6-8. Internal Partitions Associated with a Columnstore Index

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       partitions.partition_number,
       column_store_row_groups.row_group_id,
       column_store_row_groups.state_description,
       column_store_row_groups.total_rows,
       column_store_row_groups.size_in_bytes,
       column_store_row_groups.deleted_rows,
       internal_partitions.internal_object_type_desc,
              internal_partitions.rows,
       internal_partitions.data_compression_desc
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.partitions
ON partitions.partition_number = column_store_row_groups.partition_number
AND partitions.index_id = indexes.index_id
AND partitions.object_id = tables.object_id
LEFT JOIN sys.internal_partitions
ON internal_partitions.object_id = tables.object_id
WHERE tables.name = 'Sale_CCI'
ORDER BY indexes.index_id, column_store_row_groups.row_group_id;

93
Chapter 6 Columnstore Metadata

The results of this query provide at least one row per rowgroup per partition,
depending on the contents. Figure 6-7 contains a sample of this output.

Figure 6-7. Sample of internal columnstore index objects

The internal columnstore objects shown illustrate the presence of both deleted rows
and delta store contents. The deleted_rows column contains an individual count for each
rowgroup, whereas the rows column contains aggregate details.
The following are short descriptions of each new column in the result set and how it
can be used to track the growth and usage of a columnstore index.

I nternal_object_type_desc
This indicates what type of internal object is referenced, and may be one of the
following values:

• COLUMN_STORE_DELETE_BITMAP: This is the portion of the

delete bitmap associated with each rowgroup in the columnstore
index. The value in the rows column indicates how many deleted
rows are accounted for in this portion of the delete bitmap, if any.
Delete bitmaps are stored in rowstore b-tree structures. Each
rowgroup will have a delete bitmap associated with it, even if the row
count indicated as deleted is zero.

• COLUMN_STORE_DELTA_STORE: This represents delta rowgroups

that are not yet in compressed segments within the columnstore
index. These entries in the view will not exist unless a delta rowgroup
exists for that rowgroup. Delta rowgroups within the delta store are
stored in rowstore b-tree structures.

94
Chapter 6 Columnstore Metadata

• COLUMN_STORE_DELETE_BUFFER: This object is exclusive to

nonclustered columnstore indexes and buffers deleted rows from the
clustered rowstore index into the nonclustered columnstore delete
bitmap. The movement of deleted row data from the buffer into the
delete bitmap is managed via the tuple mover and can be forced via an
index reorganize command. This buffer helps improve delete speeds
on OLTP tables where update and delete operations may be frequent,
therefore helping to improve write speeds to the nonclustered
columnstore index. When data is read from the nonclustered
columnstore index, both the delete buffer and delete bitmap are read
in tandem to produce a complete picture of deleted rows in the index.
This internal object will not exist in sys.internal_partitions unless a
nonclustered columnstore index exists on a table.

• COLUMN_STORE_MAPPING_INDEX: This is used when a clustered

columnstore index also has a nonclustered rowstore index present
on it. This mapping index is used to link nonclustered index key
columns back to the correct rows within each rowgroup in the
clustered columnstore index. The mapping index is only used when
rows move between rowgroups, such as when a delta store rowgroup
is moved into a compressed rowgroup or when the tuple mover
merges rowgroups together. This internal object will not display in
sys.internal_partitions unless these conditions are met and rows are
present in the object.

95
CHAPTER 7

Batch Execution
Processing of rows in SQL Server is traditionally managed one row at a time. For
transactional workloads, this is a sensible convention, as row counts for read and write
operations are typically small.
Analytic queries that routinely operate on thousands or millions of rows do not
benefit from reading rows in this fashion. Batch mode execution is a SQL Server
execution mode that allows groups of rows to be read and passed between execution
plan operators, ultimately improving performance.
As a query is executed and each operator is processed, rows are passed between
those operators in quantities determined by the execution mode. In row mode, each
row is passed from operator to operator sequentially. In batch mode, groups of rows are
passed between operators. The result of this convention is that in batch mode, control is
passed between operators less often as rows can be handed off in fewer batches.

Row Mode Execution

Queries with a low estimated cardinality will typically use row mode execution. In row
mode execution, rows are read one by one as they are passed through each operator in
an execution plan. This sounds inefficient, but it is important to note that processing
rows in batches incurs overhead to do so. Therefore, SQL Server will make decisions at
runtime to determine if a query would be more efficiently processed in row mode or
batch mode.
Consider the query in Listing 7-1.

Listing 7-1. Transactional Query Returning a Single Row

SELECT
Employee,
[WWI Employee ID],
97
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_7
Chapter 7 Batch Execution

[Preferred Name],
[Is Salesperson]
FROM Dimension.Employee
WHERE [Employee Key] = 17;

The result of this query is a single row, as it seeks a single ID value by the primary key
in the table. Turning on the actual execution plan, the details for the clustered index seek
can be viewed, as seen in Figure 7-1.

Figure 7-1. Execution plan with details displayed

Note the execution modes provided in the query operator details. Row mode is
indicated for both the estimated and actual execution modes. Given that the underlying
table is a rowstore table and that the query only returns a single row, execution via row
mode is expected for this operation.
While columnstore indexes are optimized for analytic queries that operate against
large row counts and will usually take advantage of batch mode processing, row mode can
be chosen as the execution mode if the optimizer believes that is the most efficient option.
The query in Listing 7-2 shows a narrow query executed against a columnstore
indexed table that also happens to have a nonclustered rowstore primary key.

98
Chapter 7 Batch Execution

Listing 7-2. Query That Uses Row Mode Execution Against a Columnstore Index

SELECT
*
FROM fact.Sale
WHERE [Invoice Date Key] = '1/1/2016'
AND [Sale Key] = 198840;

This query returns a single row against a clustered columnstore index in which row
mode was used, as shown in Figure 7-2.

Figure 7-2. Execution plan with row mode execution against a columnstore index

While the execution plan operator indicates it is using columnstore storage, the
optimizer chooses row mode as the execution mode. This should not be seen as
unusual or suboptimal. Because the query only returns a single row, the use of row
mode is optimized for that expected outcome, even if the table is stored in a clustered
columnstore index.

99
Chapter 7 Batch Execution

Batch Mode Execution

Queries that have a high cardinality can benefit from batch mode execution. In this
execution mode, rows are divided into batches and processed in those batches for each
operator in an execution plan where batch mode is chosen. In columnstore indexes, this
will be the expected execution mode. For a large analytic query to not use batch mode
would be considered an outlier. If that outlier was identified as a poorly performing query,
then further investigation as to why batch mode was not used would be worthwhile.
Consider the query in Listing 7-3.

Listing 7-3. Query That Uses Batch Mode Execution Against a

Columnstore Index
SELECT
       COUNT(*),
       MIN([Invoice Date Key]),
       MAX([Invoice Date Key])
FROM fact.Sale
WHERE [Invoice Date Key] >= '1/1/2016';

This query returns only a single row, but processes many rows to crunch these
metrics. The execution plan in Figure 7-3 shows the details for this query.

Figure 7-3. Execution plan with batch mode execution on a columnstore index

100
Chapter 7 Batch Execution

Batch is the expected execution mode and is what SQL Server chose for this query.
Note that the plan details indicate that the “Actual Number of Locally Aggregated Rows”
is 29,458. This is the number of rows required by SQL Server to satisfy the query, which is
confirmed in the results shown in Figure 7-4.

Figure 7-4. Results of a simple columnstore aggregation query

Whether batch mode is chosen by the query optimizer for a given plan operator
depends on the number of rows processed by that operator and not on the number of
rows returned by a query.
Execution modes are not all-or-nothing choices for a query. The query optimizer can
choose batch mode for some operators and row mode for others and will do so based on
whatever mode it determines is most efficient for each.
A common analytic query pattern is to aggregate data from a large columnstore
index and join into dimension tables to provide lookup values where needed. The query
in Listing 7-4 illustrates a classic scenario in which both dimension and fact tables are
queried together.

Listing 7-4. Query That Joins a Large Analytic Table with a Small Lookup Table

SELECT
       City.City,
       City.[State Province],
       City.Country,
       COUNT(*)
FROM Fact.Sale
INNER JOIN Dimension.City
ON City.[City Key] = Sale.[City Key]
WHERE [Invoice Date Key] >= '1/1/2016'
AND [Invoice Date Key] < '2/1/2016'
GROUP BY City.City, City.[State Province], City.Country;

The execution plan for this query is found in Figure 7-5.

101
Chapter 7 Batch Execution

Figure 7-5. Execution plan for a query against fact and dimension tables

In the execution plan is a columnstore index scan on fact.Sale, as well as a scan

against the Dimension.City lookup table. Figure 7-6 shows the properties for each table
access operator, with the columnstore scan on the left and the rowstore scan on the right.

Figure 7-6. Operator properties for columnstore and rowstore index scans

The columnstore index scan operates in batch mode, whereas the rowstore index
scan operates in row mode. This query was executed against a database with SQL Server
2016 compatibility level (130). To test the effects of batch mode on rowstore usage,
the compatibility level is adjusted to 150 (SQL Server 2019), as shown by the query in
Listing 7-5.

102
Chapter 7 Batch Execution

Listing 7-5. Query to Alter the Database Compatibility Level to SQL Server 2019

ALTER DATABASE WideWorldImportersDW SET COMPATIBILITY_LEVEL = 150;

After this change, the query in Listing 7-4 is executed again, with the execution plan
shown in Figure 7-7.

Figure 7-7. Execution plan for a query in compatibility level 150

When the compatibility level is switched to 150, multiple new features in SQL Server
2019 are enabled. The execution plan shows the use of adaptive joins, where the join to
Dimension.City is given the option of either a clustered index scan or a clustered index
seek. The actual row counts for each operator confirm that SQL Server chose to use the
clustered index scan, due to the large number of rows returned. Figure 7-8 shows the
operator properties for the clustered index scan against Dimension.City.

103
Chapter 7 Batch Execution

Figure 7-8. Operator properties for the rowstore clustered index scan
With SQL Server 2019 features available, batch mode is chosen as the execution
mode for the rowstore clustered index scan. For a query scanning many rows, this makes
perfect sense.
Starting in SQL Server 2019, batch mode can be used by SQL Server against rowstore
tables, but will be limited to scenarios when row counts are higher and using batch
mode is expected to improve performance. Prior to SQL Server 2019, batch mode was
only available for columnstore indexes. This is a significant improvement as it can greatly
improve analytic query performance against rowstore tables.
This leads to an immediate question: If batch mode on rowstore tables can improve
analytic performance, are columnstore indexes still useful? If batch mode were the
only feature that made columnstore indexes highly performant, then that would be
a valid question. Batch mode on rowstore tables does not provide the other benefits
104
Chapter 7 Batch Execution

of columnstore indexes, such as columnstore compression, segment elimination, or

actionable metadata. For tables that service primarily analytic workloads, clustered
columnstore indexes will be the correct solution. For tables that service mixed OLAP
and OLTP workloads, batch mode on rowstore will improve performance and potentially
remove the need to create nonclustered columnstore indexes, assuming OLAP queries
are isolated and do not scan too many rows at once. Ultimately, testing is required
to determine the optimal indexing solution for a given application, but batch mode
on rowstore is a helpful tool that will aid in this decision-making process. For more
information on mixing rowstore and columnstore indexes, see Chapters 12 and 13.

How Does Batch Mode Work?

The goal of batch mode processing is to improve the throughput of data from its storage
location to a SQL Server’s CPUs. To accomplish this, SQL Server fundamentally changes
how data flows through execution plans from start to finish.
A query execution plan can contain one or many operators that retrieve, join, or
transform data before results can be returned. In row mode processing, rows are passed
between each of those operators one at a time. The basic unit of work in query execution
when using row mode is the row. Consider the execution plan shown in Figure 7-9.

Figure 7-9. Sample row mode execution plan with row counts

In the execution plan, a clustered index scan is retrieving 276 rows that are passed
into an aggregation that results in a single row that is in turn returned by the query. In
row mode, each of the 276 rows is individually transferred from the clustered index scan
operator to the stream aggregate operator.
When used, batch mode processes multiple rows together, storing the resulting data
structure as a set of vectors in memory. This results in a completely new way of data
being passed from operator to operator in an execution plan. Each batch can consume
up to 64 kilobytes and contain between 64 and 900 rows. The row batch can be visualized
in Figure 7-10.
105
Chapter 7 Batch Execution

Figure 7-10. Basic structure of a row batch

A row batch looks somewhat similar to the structure of a columnstore index and
provides similar benefits in memory as a query is executed. The qualifying rows vector
performs a similar function to the columnstore index’s delete bitmap in that it flags rows
that are no longer logically needed by the batch. For example, if a filter is applied to a
data set, it can be processed solely by updating the qualifying rows vector for the rows
that are to be filtered out.
Batch mode processing reduces CPU consumption by decreasing the number of
times that data is transferred between execution plan operators. For example, 5000 rows
processed via row mode processing will require at least 5000 CPU operations to move it
between plan operators. The same data set processed via batch mode might assign 500
rows each to ten batches. In this scenario, data can be passed between execution plan
operators using 5000/500 = 10 operations. These are approximations, but are illustrative
for how batch mode and row mode processing affect CPU consumption.
Another boon to batch mode processing is the ease in which it can take advantage of
parallelism. Batches can be processed in parallel, allowing a multicore CPU architecture
with available CPU capacity to process query plan operators more efficiently. Parallelism
is a process that requires some CPU overhead to utilize. Therefore, the amount of rows
involved in parallelism needs to be significant enough that the benefits of breaking a
workload into smaller chunks outweigh the cost to do so. Row mode does not as easily
benefit from parallelism in this regard as the effort to split single row operations into
separate parallel operations and then combine them back together is far greater than the
trivial savings afforded by that process.

106
Chapter 7 Batch Execution

While batch mode may not always be chosen for rowstore tables, it will be the default
choice for columnstore indexes. Since a rowgroup will contain up to 220 rows, the volume
of data that needs to be processed by a columnstore index scan will be large enough to
ensure that SQL Server chooses batch mode as the likely candidate to provide the best
performance for queries against that data.

Batch Mode vs. Row Mode Performance

While columnstore indexes will most often default to using batch mode processing,
rowstore tables will have a choice as to which mode to utilize. That decision will be
based on the cardinality of the query, the amount of data that needs to be read, and the
version of SQL Server involved.
There is value in directly comparing the performance of a query when it uses
row mode vs. using batch mode and measuring the performance gains provided by
batch mode.
The following query performs a simple analytic operation against a columnstore index
twice: once using compatibility level 120 (SQL Server 2014) and once using compatibility
level 130 (SQL Server 2016). Listing 7-6 provides the code to perform this comparison.

Listing 7-6. Analytic Query Executed in Two Compatibility Levels

ALTER DATABASE WideWorldImportersDW SET COMPATIBILITY_LEVEL = 120;

SELECT
Sale.[City Key],
COUNT(*)
FROM fact.Sale
GROUP BY Sale.[City Key]

ALTER DATABASE WideWorldImportersDW SET COMPATIBILITY_LEVEL = 130;

SELECT
Sale.[City Key],
COUNT(*)
FROM fact.Sale
GROUP BY Sale.[City Key]
107
Chapter 7 Batch Execution

The execution plans that result from these sample queries provide an immediate
clue that a significant performance difference occurred here, as seen in Figure 7-11.

Figure 7-11. Execution plans showing significant performance discrepancies

The two execution plans appear nearly identical, except that the query cost for the
first is 90%, whereas the second query is 10%. In addition, aggregate pushdown occurred,
allowing the GROUP BY in our query to be processed in line with the columnstore index
scan. This prevented the need to push all 228265 rows one at a time on to the hash match
operation. Figure 7-12 compares the details for each columnstore index scan.

108
Chapter 7 Batch Execution

Figure 7-12. Execution plan details for row mode vs. batch mode operation
There are a handful of notable differences between the expensive execution plan on
the left and the efficient execution plan on the right:
• Batch mode was used successfully to process 228265 rows in the
second plan.
• CPU consumption using batch mode is nearly ten times less than
using row mode.
• Actual Number of Locally Aggregated Rows documents aggregate
pushdown.
Making use of batch mode allows SQL Server to also take advantage of aggregate
pushdown. The combination of these features was what allowed for CPU to be so greatly
reduced.
This raises an important aspect of batch mode processing that further increases its
effectiveness on analytic workloads: batch mode enables other powerful performance-
improving features to be used when a query is optimized. Some of these features (along
with earliest SQL Server versions they were enabled) include
• Adaptive joins (SQL Server 2017)
• Memory grant feedback (SQL Server 2017)
• Aggregate pushdown (SQL Server 2016)
109
Chapter 7 Batch Execution

Generally speaking, when SQL Server chooses to use batch mode, performance
will be improved with its usage. Similarly, when additional intelligent query processing
features are used (such as adaptive joins or aggregate pushdown), they will also have a
positive impact on performance.
Testing batch mode vs. row mode processing can be challenging as forcing a query to
use one over the other is not always straightforward. One way to make this testing easier
is to use database scoped configuration changes to temporarily disable these features.
The T-SQL in Listing 7-7 provides the syntax needed to disable batch mode on rowstore,
as well as batch mode memory grant feedback and batch mode adaptive joins.

Listing 7-7. Query That Disables Optimization Features (for Testing

Purposes Only)
ALTER DATABASE SCOPED CONFIGURATION SET BATCH_MODE_ON_ROWSTORE = OFF;

ALTER DATABASE SCOPED CONFIGURATION SET BATCH_MODE_MEMORY_GRANT_

FEEDBACK = OFF;

ALTER DATABASE SCOPED CONFIGURATION SET BATCH_MODE_ADAPTIVE_JOINS = OFF;

Note that these features should only be disabled for testing purposes only and
should not be turned off in a production environment unless there are exceptional and
well-documented reasons to do so.
Compatibility levels may also be adjusted for testing and demonstration purposes.
This can help model the impact of an upgrade or to allow a SQL Server upgrade
to functionally occur at a slower pace over time. By incrementally adjusting the
compatibility level up by one level at a time from the original SQL Server version to the
new version, risk can be isolated and mitigated in steps, rather than all at once. This
process also provides a rollback mechanism, if needed, since compatibility modes can
be lowered, in the event of an emergency.

110
CHAPTER 8

Bulk Loading Data

Any analytic data store requires the ability to perform data loads quickly and efficiently.
Bulk loading is a reduced logging process that allows data to be inserted directly into a
columnstore index. This not only bypasses the delta store, but results in a transaction
size that reflects the compression of the target data, greatly reducing the amount of data
written to the transaction log when this process is used.

Bulk Load Processes Explained

Traditional transactional workloads are fully logged. When fully logged, enough
metadata is written to the transaction log so that SQL Server can recover from a failed
transaction. In addition, data is written to the transaction log to ensure that point-in-
time recovery is possible.
Fully logged transactions in OLTP scenarios are ideal as they allow a database to be
rolled back to a specific point in time in the event of application problems. In addition,
point-in-time recovery allows for research and forensics into the details of transactions,
when needed.
Sometimes, transactional writes are intentionally subdivided into smaller batches to
ensure that each transaction is short and fast and causes minimal contention with other
queries against the same data.
The cost of fully logged transactions is that more data needs to be written to the
transaction log, which results in

• More storage space consumed in the transaction log

• More storage space consumed in transaction log backup files

• More CPU/memory to process transaction log backups

• Longer query duration for write operations

111
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_8
Chapter 8 Bulk Loading Data

Analytic workloads differ greatly in their recovery needs as data loads tend to
be larger, less frequent, and asynchronous. When analytic data loads fail, the most
common recourse is to investigate the problem, resolve it, and then rerun the data load.
Point-in-time recovery within data load processes is less important than simply having
recovery available to a point in time prior to the data load. Therefore, OLAP data loads
can benefit greatly from minimally logged insert operations.
Outside of columnstore indexes, bulk loading data is limited to a handful of write
operations, such as
• BCP

• Bulk insert operations

• Select into operations

• Partition switching

Because point-in-time recovery can be critical to transactional systems, any process

that utilizes bulk loading needs to be documented well enough to not accidentally be
used in a scenario where it is not desired.
Columnstore indexes use built-in bulk load processes that automatically insert larger
batches directly into rowgroups without using the delta store. This greatly improves
insert speeds for large analytic data loads and reduces transaction log bloat resulting
from those processes.

Bulk Loading into Columnstore Indexes

The breakpoint for deciding whether to use a bulk load process or to use the delta store
is 102,400 rows. Inserts of less than 102,400 rows will always be written to the delta store
via a fully logged process, whereas inserts of 102,400 rows or more will be written directly
to rowgroups in the columnstore index, bypassing the delta store. This decision can be
shown in Figure 8-1.

112
Chapter 8 Bulk Loading Data

Figure 8-1. Determination of process for bulk loading data into a

columnstore index

Unlike some of the other types of minimally logged insert operations in SQL Server,
there are no prerequisites to take advantage of bulk loading data into a columnstore index.
There is no need to adjust isolation levels, use explicit table locking, or adjust parallelism
settings. Since bulk loading data will be the desired operation for large inserts into
columnstore indexes, it is the default and will be used automatically when possible.
If an insert operation contains more than 220 (1,048,576) rows, it will be subdivided
into inserts in the following fashion:

1. Each batch of 220 rows will be bulk inserted into the columnstore
index until there are less than 220 rows remaining to insert.
2. If the remainder of rows is greater than or equal to 102,400, it will
also be bulk inserted into the columnstore index.

3. If the remainder of rows is less than 102,400, it will be inserted into

the delta store.
113
Chapter 8 Bulk Loading Data

For example, if an insert of 3,000,000 rows is made to a columnstore index, it will be

broken down as follows:

• 2 bulk loads of 1,048,576 rows each

• 1 bulk load of 902,848 rows (the remainder)

Alternatively, if an insert of 2,100,000 rows occurs, it will be processed like this:

• 2 bulk loads of 1,048,576 rows each

• 2,848 rows (the remainder) inserted into the delta store

SQL Server can bulk load data into a columnstore index in multiple parallel threads
so long as the data for each thread is targeting a different data file. This is automatic and
requires no particular user action to accomplish. Columnstore bulk insert operations
acquire exclusive locks against target rowgroups, and so long as parallel inserts target
separate data files, they are guaranteed to not overlap the rowgroups they are inserting into.
When data is inserted into a partitioned columnstore index, that data is first assigned
a partition and then each group of rows is inserted into their respective partitions.
Therefore, whether bulk load processes are used will be dependent on the numbers of
rows inserted into the target partition, rather than the total number of rows inserted
by the source query. Typically, analytic tables will have a current/active partition that
accepts new data, whereas the remaining partitions will contain older data that is no
longer written to (outside of one-off data loads, software releases, or maintenance
events).

erformance of Bulk Loading into

P
Columnstore Indexes
To demonstrate the impact of bulk loading data into a columnstore index, a test will be
conducted against a rowstore copy of Fact.Sale. The table’s creation is not shown here,
but is an identical copy, with only a clustered primary key and page compression, rather
than a clustered columnstore index. This insert query is shown in Listing 8-1.

114
Chapter 8 Bulk Loading Data

Listing 8-1. An Insert of 102,400 Rows into a Clustered Rowstore Index

INSERT INTO fact.Sale_Transactional

The insert takes about 1 second. With the rows inserted, an undocumented but
useful system function will be used to read the contents of the transaction log and
determine the size of the transaction, as seen in Listing 8-2.

Listing 8-2. Query to Calculate the Transaction Size for a Clustered

Rowstore Insert
SELECT
fn_dblog.allocunitname,
SUM(fn_dblog.[log record length]) AS log_size
FROM sys.fn_dblog (NULL, NULL)
WHERE fn_dblog.allocunitname = (
'Fact.Sale_Transactional.PK_Fact_Sale_Transactional')
GROUP BY fn_dblog.allocunitname;

The result is a single row that indicates the total log size for the insert transaction, as
seen in Figure 8-2.

Figure 8-2. Transaction size for an insert of 102,400 rows into a clustered
rowstore index

115
Chapter 8 Bulk Loading Data

The log size returned for the rowstore insert was 225,856 bytes, or about 220KB. The
same insert will now be performed against a clean copy of the table with a columnstore
index, as seen in Listing 8-3.

Listing 8-3. An Insert of 102,400 Rows into a Clustered Columnstore Index

INSERT INTO fact.Sale_CCI_Clean_Test

This insert takes less than a second to complete. A new clean table is used to ensure
that there are no residual transactions from previous demonstrations that would pollute
the log. The results from fn_dblog() are seen in Figure 8-3.

Figure 8-3. Transaction size for an insert of 102,400 rows into a clustered
columnstore index

Note that the transaction log size for the insert is 118,192 bytes, or about 115KB. This
constitutes a significant reduction in transaction size when compared to a page
compressed rowstore index.
With the impact that bulk loading can have on transaction size demonstrated, it is
important to illustrate the difference between inserting 102,400 rows into a columnstore
index and inserting 102,399 rows. The T-SQL in Listing 8-4 inserts 102,399 rows into a
newly created clustered columnstore index.

116
Chapter 8 Bulk Loading Data

Listing 8-4. An Insert of 102,399 Rows into a Clustered Columnstore Index

INSERT INTO fact.Sale_CCI_Clean_Test_2

This takes about a second to execute. The query to pull data from fn_dblog() needs to
be adjusted slightly to accommodate the log growth due to both the columnstore index
itself and the delta store. This is shown in Listing 8-5.

Listing 8-5. Query to Calculate Log Growth for Columnstore Index and
Delta Store
SELECT
fn_dblog.allocunitname,
SUM(fn_dblog.[log record length]) AS log_size
FROM sys.fn_dblog (NULL, NULL)
WHERE fn_dblog.allocunitname IN (
'Fact.Sale_CCI_Clean_Test_2.CCI_Sale_CCI_Clean_Test_2',
'Fact.Sale_CCI_Clean_Test_2.CCI_Sale_CCI_Clean_Test_2(Delta)')
GROUP BY fn_dblog.allocunitname;

Note that the delta store needs to be referenced separately to be included in the
results. The transaction log space consumed by each object is seen in Figure 8-4.

117
Chapter 8 Bulk Loading Data

Figure 8-4. Transaction size for an insert of 102,399 rows into a clustered
columnstore index
The insert into the delta store was significantly more expensive than any operation
demonstrated thus far in this chapter, with a transaction size of 1,300,448, or about
1.2GB. As a result, there is a significant incentive to take advantage of bulk loading data
into columnstore indexes when possible.

Trickle Insert vs. Staged Insert

While inserting into the delta store may seem expensive, it is nowhere near as costly as
repeated inserts into a columnstore index would be. If an analytic table is often targeted
with many smaller insert operations, there is significant value in collecting those rows
into a temporary table first, prior to inserting into the columnstore index.
In workloads involving rowstore tables, large insert operations may be batched into
small groups of rows in order to reduce contention and decrease the transaction size for
each insert. Columnstore indexes do not benefit from micro-batching. When migrating
code from rowstore tables to columnstore tables, resources can be saved by adjusting
data load processes to operate on significantly larger batches. Batches of 220 (1,048,576)
rows are optimal, though a batch size greater than or equal to 102,400 will ensure that
the delta store is not needed. If insert operations contain tens of millions of rows, then
those inserts can be broken down into more manageable subunits to prevent a staging or
temporary table from becoming too large.
The best practices for loading data into columnstore indexes are as follows:
1. Load at least 102,400 rows per insert operation.
2. When possible, load 1,048,576 rows per batch.
3. Only load less than 102,400 rows for the remaining rows at the end
of the data load.
While this guidance may make it seem as though the delta store should be avoided at
all costs, the goals of an analytic data store should not be compromised for this purpose.
Delaying reporting data to prevent delta store usage is not worthwhile and will not
118
Chapter 8 Bulk Loading Data

save enough resources to be meaningful. Actively avoiding the insert of repeated small
batches will ensure that insert performance is quite good. If the delta store is needed for
the tail end of a data load process, then there is no reason to avoid it.

Other Data Load Considerations

Vertipaq optimization, which can greatly improve the effectiveness of columnstore index
compression, is not used when a nonclustered rowstore index is present on a clustered
columnstore index. Over time, this will result in rowgroups that will consume more
resources than they would otherwise.
For scenarios where nonclustered rowstore indexes are required to support
analytic processes, consider the following options for managing performance on the
columnstore index.

Drop Nonclustered Indexes During Data Loads

A simple solution to ensure that Vertipaq optimization is always used is to drop
nonclustered rowstore indexes during data loads and re-create them afterward.
Building the nonclustered indexes is an online operation and will not interfere with
other queries against the table. The downside to this approach is that any analytics that
use the nonclustered indexes will be impacted during the data load process, until the
nonclustered index rebuild is complete. Additional resources are also required to handle
the nonclustered index build. This approach is most effective when the nonclustered
indexes are small and relatively quick to build.
On partitioned tables, only the partitions with data that has changed require this
consideration. Typically, partitions containing older data that is primarily static and
unchanging in nature (not hot data) can be ignored.
For scenarios when nonclustered indexes are necessary and cannot be dropped,
consider periodic maintenance to address inefficient compression over time. If a
quarterly maintenance can be scheduled, then the following actions can be taken:

1. Drop nonclustered indexes on active partitions.

2. Rebuild the clustered columnstore index on active partitions.

3. Re-create nonclustered indexes.

119
Chapter 8 Bulk Loading Data

As a planned maintenance event, these are relatively innocuous tasks that should
not consume an unusual amount of time, even on a larger table.

Columnstore Reorganize Operations with Each Data Load

When a columnstore index is read, its contents plus the delta store are read together
to produce the necessary output. While the delta store cannot get excessively large,
columnstore indexes with periodic data loads can improve read speeds slightly by
forcefully compressing the contents of the delta store into the columnstore index. The
query in Listing 8-6 examines the contents of the small columnstore index created
earlier.

Listing 8-6. Query to Return Details About the Structure of a Columnstore Index

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       partitions.partition_number,
       column_store_row_groups.row_group_id,
       column_store_row_groups.state_description,
       column_store_row_groups.total_rows,
       column_store_row_groups.size_in_bytes,
       column_store_row_groups.deleted_rows,
       internal_partitions.internal_object_type_desc,
       internal_partitions.rows,
       internal_partitions.data_compression_desc
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.partitions
ON partitions.partition_number = column_store_row_groups.partition_number
AND partitions.index_id = indexes.index_id
AND partitions.object_id = tables.object_id

120
Chapter 8 Bulk Loading Data

LEFT JOIN sys.internal_partitions

ON internal_partitions.object_id = tables.object_id
WHERE tables.name = 'Sale_CCI_Clean_Test_2'
ORDER BY indexes.index_id, column_store_row_groups.row_group_id;

The results are found in Figure 8-5.

Figure 8-5. Rowgroup information for a small columnstore index

Note that the entire contents of the columnstore index (102,399 rows) reside in the
delta store. The delete bitmap exists as a default and is currently empty. If the operator
wishes to move the contents of the delta store into columnstore rowgroups, this can be
accomplished by the index maintenance command in Listing 8-7.

Listing 8-7. Index Reorganize Command to Compress Delta Store

ALTER INDEX CCI_Sale_CCI_Clean_Test_2 ON Fact.Sale_CCI_Clean_Test_2

REORGANIZE WITH (COMPRESS_ALL_ROW_GROUPS = ON);

Once complete, the delta store would be compressed and ready to move into the
columnstore index, as seen in Figure 8-6.

Figure 8-6. Effects of index maintenance on a columnstore delta rowgroup

The results show that the delta rowgroup affected by the index maintenance is
now in an intermediary state. A new object has been created and compressed with
the contents of the delta rowgroup inserted into it. The previous delta rowgroup (an
uncompressed heap) is left in a tombstone state for the tuple mover to remove at a
later point in time. Note the significant size difference between an uncompressed delta
rowgroup and a compressed rowgroup.

121
Chapter 8 Bulk Loading Data

At this time, running the same ALTER INDEX statement as before will force the
tuple mover to complete this cleanup. Alternatively, waiting for a short amount of time
to pass will achieve the same results. After 5 minutes have passed, the contents of this
columnstore index are as seen in Figure 8-7.

Figure 8-7. Columnstore index contents after the tuple mover executes

Once the tuple mover completes its cleanup process, all that remains is a single
compressed columnstore rowgroup.
Performing maintenance like this is not necessary but can improve query read
speeds against a columnstore index after a data load is completed. Index maintenance
will be discussed in far more detail in Chapter 14, including its use as part of data loads
and other columnstore processes.

S
ummary
When loading data into a columnstore index, bulk loading ensures the fastest possible
load speeds while minimizing server resource consumption. Bulk loading large numbers
of rows in fewer batches is far more efficient than using trickle or small batch inserts.
By focusing data load processing around maximizing the use of compressed
columnstore rowgroups and minimizing delta store usage, overall columnstore index
performance can be improved, both for the data load processes and analytics that use
the newly loaded data.

122
CHAPTER 9

Delete and Update

Operations
Modifying data within a highly compressed structure is expensive and requires
additional processes to manage it successfully. Whereas insert operations can benefit
from bulk loading to streamline data load processes, delete and update operations
require using the delete bitmap and delta store to manage changes to existing data.

The Cost of Modifying Data

In Chapter 5, columnstore compression was discussed in extensive detail. An important
takeaway is that some forms of compression, such as run-length encoding, depend on
the exact contents of the underlying data. Columnstore compression is exceptionally
good at shrinking data by a considerable amount by applying a variety of encoding and
compression algorithms and optimizations. Consider the sample data in Figure 9-1.

Figure 9-1. Sample data encoded with run-length encoding

123
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_9
Chapter 9 Delete and Update Operations

The sample data is encoded with dictionary compression, reordered with Vertipaq
optimization, and further compressed with run-length encoding. If a process deletes the
last two rows in the table, then it is necessary to decompress the data fully, delete the
rows, and then recompress it. Figure 9-2 shows how this process would impact the data.

Figure 9-2. Deleting rows from run-length encoded data

The resulting set of indexed ID groups contains only 4 rows, rather than 5 (as index
ID 4 has been removed), and the count for the index ID 3 has been reduced.
If the table were larger and the deletion impacted more rows, then it is likely that
many rowgroups would need to be decompressed, adjusted, and recompressed. In the
process, many pages would need to be updated. This operation would quickly grow to be
prohibitively expensive. A balance needs to be maintained between the speed of write
operations and the speed of read operations, and in this scenario, the ability to load and
modify data quickly needs to be prioritized.

124
Chapter 9 Delete and Update Operations

D
elete Operations
In a columnstore index, the cost to decompress rowgroups, delete rows, and recompress
them is prohibitively high. The more rowgroups a delete operation targets, the greater
this cost would become. To mitigate this cost, a structure called the delete bitmap is used
to track deletions from the columnstore index.
The delete bitmap is a heap that references the rows within underlying rowgroups.
When a row is deleted, the data within the rowgroup remains unchanged and the
delete bitmap is updated with a reference to the deleted row. As such, deleted rows in
columnstore indexes can be seen as soft deleted, where removed rows are flagged, but
not physically deleted.
Note that delete operations against rows in the delta store do not need to use the
delete bitmap as the delta store is a rowstore heap and rows can simply be deleted from
it as needed, without the need for soft deletes.
When a query is executed against a rowgroup containing deleted rows, the delete
bitmap is consulted and deleted rows are excluded from the results. The delete bitmap
may be visualized as seen in Figure 9-3.

Figure 9-3. The delete bitmap and its relationship to a columnstore index

The underlying rows within the rowgroup are unchanged when a deletion occurs
against them. Instead, the delete bitmap tracks which rows are deleted and is consulted
when future queries are issued against this rowgroup. Because of this, deleting data in a
columnstore index will not reclaim storage space.

125
Chapter 9 Delete and Update Operations

Consider the query shown in Listing 9-1.

Listing 9-1. Query That Deletes Data from a Columnstore Index

DELETE
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] = '1/1/2016';

When executed, this query will delete all sale data that was invoiced on 1/1/2016.
Before doing so, the rowgroup metadata can be consulted to confirm that there
are currently no deleted rows in the rowgroup. The query in Listing 9-2 will return
metadata about rowgroups in this columnstore index, including the number of deleted
rows, if any.

Listing 9-2. Query That Returns Metadata About Deleted Rows in Rowgroups

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       partitions.partition_number,
       column_store_row_groups.state_description,
       column_store_row_groups.total_rows,
       column_store_row_groups.size_in_bytes,
       column_store_row_groups.deleted_rows
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.partitions
ON partitions.partition_number = column_store_row_groups.partition_number
AND partitions.index_id = indexes.index_id
AND partitions.object_id = tables.object_id
WHERE tables.name = 'Sale_CCI'
ORDER BY indexes.index_id, column_store_row_groups.row_group_id;

126
Chapter 9 Delete and Update Operations

The rowgroup metadata can be seen in Figure 9-4.

Figure 9-4. Rowgroup metadata for a columnstore index, including deleted rows

As is expected of a columnstore index that has had no delete operations executed

against it, the deleted_rows column in sys.column_store_row_groups shows zero for each
rowgroup. Executing the query in Listing 9-1 will delete a total of 15,400 rows from the
columnstore index. Figure 9-5 shows the metadata details for the columnstore index
after the deletion has completed.

Figure 9-5. Rowgroup metadata after a deletion has occurred

Each rowgroup now contains some number of deleted rows, depending on the
number of rows within them that happen to have been invoiced on 1/1/2016. Note that
the total rows in each rowgroup and the size in bytes have not changed. This is expected
and reflects the fact that rows were soft deleted via the delete bitmap.

127
Chapter 9 Delete and Update Operations

Because rows were removed from all rowgroups in the index, a deletion without the
aid of the delete bitmap would have required a rebuild of the entire index, which would
be a prohibitively expensive operation to perform while a process waits for a single day’s
worth of rows to be deleted. Instead, the delete of 15,400 rows completed exceptionally
quickly, in under a second, thanks to the delete bitmap!
The only way to clean up deleted rows and free up the space consumed within
rowgroups is to perform an index rebuild on the columnstore index or on any partitions
impacted by the deletion. This is generally not needed unless the amount of deleted data
becomes large with respect to the overall size of the index. Like with classic rowstore
indexes, fragmentation is not problematic until there is too much of it, at which time a
rebuild can resolve it. Note that index reorganize operations do not remove deleted rows
from columnstore indexes! Chapter 14 dives into index maintenance in detail, discussing
how and when it is needed, and best practices for its use.

U
pdate Operations
In a columnstore index, an update is executed as two operations: a delete and an insert.
While logically, the update will perform as a single atomic unit, under the covers it will
consist of

1. A set of deleted rows, flagged in the delete bitmap

2. A set of newly inserted rows, written to the delta store

This means that an update operation will need to write to both the delete bitmap
and the delta store in order to complete successfully. It is also important to note that the
insert operations that result from an update operation against a columnstore index will
exclusively use the delta store and cannot take advantage of bulk load processes.
Before continuing, a rebuild will be executed against the columnstore index to allow
for easier visualization of the results. The query in Listing 9-3 will rebuild the index.

Listing 9-3. Query to Rebuild a Columnstore Index, Removing the Delete Bitmap

ALTER INDEX CCI_fact_sale_CCI ON Fact.Sale_CCI REBUILD;

With the data now clean, the query in Listing 9-4 can be used to return metadata
about all rowgroups in the columnstore index, including both the delta store and
delete bitmap.

128
Chapter 9 Delete and Update Operations

Listing 9-4. Query to Return Delete Bitmap and Delta Store Metadata for
Rowgroups in a Columnstore Index
SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       partitions.partition_number,
       column_store_row_groups.row_group_id,
       column_store_row_groups.state_description,
       column_store_row_groups.total_rows,
       column_store_row_groups.size_in_bytes,
       column_store_row_groups.deleted_rows,
       internal_partitions.internal_object_type_desc,
       internal_partitions.rows
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.partitions
ON partitions.partition_number = column_store_row_groups.partition_number
AND partitions.index_id = indexes.index_id
AND partitions.object_id = tables.object_id
LEFT JOIN sys.internal_partitions
ON internal_partitions.object_id = tables.object_id
WHERE tables.name = 'Sale_CCI'
ORDER BY indexes.index_id, column_store_row_groups.row_group_id;

The results in Figure 9-6 show a clean columnstore index with no deleted rows or
entries in the delta store.

129
Chapter 9 Delete and Update Operations

Figure 9-6. Metadata for rowgroups with no deletes/updates against them

With a pristine columnstore index available, the effects of update operations can be
easily visualized against it. Consider the query shown in Listing 9-5.

Listing 9-5. Query to Identify a Data Set to Update

SELECT
*
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] = '1/2/2016';

This SELECT query identifies a total of 18,150 rows in the table that match the filter
criteria of an invoice date of 1/2/2016. Next, an update will be made against two columns
in the table, as shown in Listing 9-6.

Listing 9-6. Query to Update Data from a Columnstore Index

UPDATE Sale_CCI
SET [Total Dry Items] = [Total Dry Items] - 1,
[Total Chiller Items] = [Total Chiller Items] + 1
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] = '1/2/2016';

Returning to the rowgroup metadata query in Listing 9-4, the results of the UPDATE
statement can be reviewed, with a sample seen in Figure 9-7.

130
Chapter 9 Delete and Update Operations

Figure 9-7. Metadata for rowgroups after updating 18150 rows

The metadata after the UPDATE was executed shows that every rowgroup in the
columnstore index was impacted as 18,150 rows were deleted and 18,150 rows were
then inserted. Sys.internal_partitions shows a delete bitmap and delta store object,
each containing 18,150 total rows. The rowgroup detail illustrates how many rows were
updated per rowgroup. In addition, the new rowgroup (number 25) shows the new open
delta store that was created for the newly inserted rows.
The resulting structure underscores the fact that an UPDATE against a columnstore
index executes as a combination of discrete delete and insert operations. Performing
each of those operations sequentially would yield similar results.
Consider an update to all rows in this columnstore index for the range of invoice
dates from 1/3/2016 up to 1/8/2016. A count of these rows shows a total of 148,170
that match that date filter. Before running an update, the columnstore index will be
rebuilt, which will clean up the deleted rows and delta stores, allowing for a better
demonstration. Listing 9-7 provides the query to rebuild the index and then update
these rows.

Listing 9-7. Query to Update 148,170 Rows in a Columnstore Index After

Rebuilding the Index
ALTER INDEX CCI_fact_sale_CCI ON Fact.Sale_CCI REBUILD;
GO

UPDATE Sale_CCI
SET [Total Dry Items] = [Total Dry Items] - 1,
[Total Chiller Items] = [Total Chiller Items] + 1
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/3/2016'
AND [Invoice Date Key] < '1/8/2016';
131
Chapter 9 Delete and Update Operations

When updating 148,170 rows, the first thing to note is that it took 5 full seconds to
execute! The previous update of 18,150 rows completed almost instantly after being
executed. Viewing the metadata reveals the reason for this, as seen in Figure 9-8.

Figure 9-8. Metadata for rowgroups after updating 148,170 rows

The key takeaway from the metadata following the larger update is that the delta
store contains all 148,170 rows updated in the operation. Normally, any INSERT
operations of 102,400 rows or more will benefit from a minimally logged bulk insert
process, but UPDATE operations cannot benefit from it. Because the UPDATE is
composed of both an INSERT and DELETE in the same transaction, it is not possible to
fork the single transaction into a fully logged DELETE and a minimally logged INSERT.
Bulk insert processes on columnstore indexes can save immense resources when
used, but are not allowed to violate the ACID (Atomic, Consistent, Isolated, and Durable)
properties of a SQL Server database. Attempting to splice together a fully logged and
minimally logged transaction into a single larger transaction would require the creation
of a transaction that would provide point-in-time restore capabilities for the DELETE,
but not the INSERT. While it is possible to conceive of ways for SQL Server to architect its
way around that limitation, doing so would be confusing to anyone using a columnstore
index and could result in unexpected results when restoring databases containing
columnstore indexes that are subject to frequent UPDATE operations.
The key takeaway is that an UPDATE against a columnstore index will be comprised
of a DELETE operation and a fully logged INSERT into the delta store. As the number of
rows updated increases, the performance of those operations will decrease dramatically.
The delta store was built to handle small numbers of rows – either trickle loads or the
residual rows from a larger data load process. It was not built for large volumes of rows at
one time and as a matter of course will perform poorly for those applications.

132
Chapter 9 Delete and Update Operations

There is one final demonstration to underscore this challenge, and that is to

update more rows than would fit into a single rowgroup. Consider the sample query in
Listing 9-8.

Listing 9-8. Query to Count Rows with a Wide Date Range on a

Columnstore Index
SELECT
COUNT(*)
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/8/2016'
AND [Invoice Date Key] < '3/5/2016';

When executed, the result is 1,159,180 rows counted. This is greater than the
1,048,576 rows that can be stored in a single columnstore rowgroup. Listing 9-9 rebuilds
the index once more and then performs an update against all 1,159,180 rows identified
in Listing 9-9. This single UPDATE operation will result in the deletion of old rows, the
insertion of new rows into the delta store, and the compression of most of the new rows
into rowgroups.

Listing 9-9. Query to Update 1,159,180 Rows in a Columnstore Index After

Rebuilding the Index
ALTER INDEX CCI_fact_sale_CCI ON Fact.Sale_CCI REBUILD;
GO

UPDATE Sale_CCI
SET [Total Dry Items] = [Total Dry Items] - 1,
[Total Chiller Items] = [Total Chiller Items] + 1
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/8/2016'
AND [Invoice Date Key] < '3/5/2016';

The update of 1,159,180 rows took almost a minute to complete. Figure 9-9 shows
the resulting columnstore metadata immediately after this large UPDATE operation
completes.

133
Chapter 9 Delete and Update Operations

Figure 9-9. Metadata for rowgroups after updating 1,159,180 rows

Note that there are multiple delta stores present. The open delta store (number 26)
will remain open to accept future inserted data. The closed delta store will be processed
by the tuple mover asynchronously and be compressed into a permanent columnstore
rowgroup. Figure 9-10 shows the metadata for this columnstore index after a minute has
passed and the tuple mover has executed.

Figure 9-10. Metadata for rowgroups after updating 1,159,180 rows and allowing
the tuple mover to execute

The rowgroup labeled in a state of “TOMBSTONE” (number 25) is no longer a logical

part of the columnstore index and will be cleaned up in the future, either via the tuple
mover or index maintenance. Rowgroup number 27 contains the contents of rowgroup
number 25 when it was processed by the tuple mover. To summarize the changes made
to the columnstore index by the update of 1,159,180 rows:

1. 1,159,180 rows were marked as deleted in the delete bitmap across

all rowgroups.

2. 1,048,576 rows were inserted into one delta store, filling it to

capacity.

134
Chapter 9 Delete and Update Operations

3. 110,604 rows were inserted into another delta store and remain in
an open state awaiting future inserted rows.

4. The 1,048,576 rows in the full delta store are processed by the
tuple mover and compressed into a permanent rowgroup.

This is a significant amount of work for a single UPDATE statement, and it scales
poorly with increasing row counts.
In general, updates should be limited to small row counts or avoided altogether.
There is value in rethinking how code is written to refactor an UPDATE against
a columnstore index into a set of deletes and inserts or to manage an update via
intermediary processes that avoid it altogether. The performance of UPDATE operations
on columnstore indexes will be unpredictable and a large enough row count will result
in transactions large enough to create resource pressure on the SQL Server.
Chapter 15 (best practices) will discuss in greater detail how to avoid updates and a
variety of tactics that can be used when migrating UPDATE code from rowstore indexes
into columnstore indexes.

135
CHAPTER 10

Segment and Rowgroup

Elimination
A key component to how columnstore indexes function is that every segment within
the index is its own building block. Each may be read individually or in groups, but the
number of segments read via any query can be reduced by efficient architecture and
optimal query patterns. Reducing segments read directly reduces IO, improves query
speed and memory-related performance metrics, such as page life expectancy.

S
egment Elimination
Each segment represents data for a single column over a set of rows. When a query is
executed against a columnstore index, SQL Server needs to determine which rowgroups
are required to return the requested result set. Within those rowgroups, only the
segments containing data for the selected columns will be read. Therefore, queries that
select less columns will read fewer segments, thereby reducing IO, memory usage, and
query runtimes.

137
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_10
Chapter 10 Segment and Rowgroup Elimination

Consider the columnstore index shown in Figure 10-1.

Figure 10-1. Sample columnstore index with six rowgroups and eight columns

With 6 rowgroups and 8 columns, this columnstore index contains a total of

48 segments. Because each column is broken into its own set of segments (one per
rowgroup), it is not necessary to read segments for columns that are not needed for a
query. If a sample query were to only read columns A and B, the result would be that the
remaining segments can be automatically skipped, as shown in Figure 10-2.

138
Chapter 10 Segment and Rowgroup Elimination

Figure 10-2. Impact of segment elimination on a query for two columns

By querying for only two of the possible eight columns, the number of segments
read was reduced by 75%. This convention holds true for columnstore indexes with any
number of columns. If this table had 24 columns, then reading only 2 of them would
mean that 22 columns would not be read. The result of this feature is that the number
of segments read in a columnstore index query will be proportional to the number of
columns involved in the query.

139
Chapter 10 Segment and Rowgroup Elimination

The columns needed to satisfy a query also include any referenced by the WHERE
clause, as well as in aggregates (GROUP BY/HAVING). If a query includes views
or functions, then their contents will be evaluated to determine which columns
are required to execute a query. Listing 10-1 provides an example query against a
columnstore index.

Listing 10-1. Sample Columnstore Query Using a Filter and Aggregate

SELECT
[City Key],
COUNT(*)
FROM fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/1/2016'
GROUP BY [City Key]
ORDER BY COUNT(*) DESC;

This analytic query calculates a count of sales per city for a given time period. Of the
21 columns in Sale_CCI, only 2 were required to complete this calculation: City Key and
Invoice Date Key. As a result, 19 of the 21 column’s worth of segments will be ignored
when executing this query.
Alternatively, rowstore tables store rows sequentially with each column for each row
stored together on pages. While reading less columns in rowstore indexes can reduce
the amount of data presented to an application, it does not reduce the quantity of pages
read into memory in order to retrieve the specific columns for a query. Therefore,
reducing the number of columns queried in a columnstore index will provide immediate
performance benefits that are not as pronounced in rowstore indexes.
Segment elimination is a simple and powerful tool that can be simplified into a
single optimization technique: Write queries to only include the required columns.
Since rowgroups can contain up to 220 rows, the cost of querying unnecessary columns
can be quite high. Whereas transactional queries against rowstore tables often operate
on small numbers of rows, analytic queries can access millions of rows at a time.
Therefore, the perceived convenience of SELECT * queries will hinder performance on a
columnstore index.
Consider the query in Listing 10-2.

140
Chapter 10 Segment and Rowgroup Elimination

Listing 10-2. A SELECT * Query Against a Columnstore Index

SELECT
*
FROM fact.Sale_CCI
WHERE [Invoice Date Key] = '2/17/2016';

This SELECT * query returns all columns for sales on a specific date. While only
30,140 rows are returned, all 21 columns need to have their segments retrieved as part of
the operation. Figure 10-3 provides the output of STATISTICS IO for this query.

Figure 10-3. IO for a sample SELECT * query

Note that a total of 68,030 LOB logical reads are recorded. Listing 10-3 contains an
alternative query in which only a small set of columns required by an application are
returned.

Listing 10-3. A Query Against a Columnstore Index Requesting Only

Three Columns
SELECT
       [Sale Key],
       [City Key],
       [Invoice Date Key]
FROM fact.Sale_CCI
WHERE [Invoice Date Key] = '2/17/2016';

Here, only three of the columns are requested, instead of all of them. The resulting
IO can be viewed in Figure 10-4.

Figure 10-4. IO for a sample query requesting data from three columns

141
Chapter 10 Segment and Rowgroup Elimination

When the column list was reduced to only those needed for the query, LOB logical
reads were reduced from 68,030 to 31,689. The performance improvement for omitting
columns will vary based on the data type, compression, and contents of each column.
Omitting a text column with no repeated values will reduce IO far more than omitting a
BIT column.
Segment elimination is an easy way to improve query speeds while also reducing
resource consumption. Anyone writing queries against columnstore indexes should
consider which columns are required for their workloads and ensure that no extra
columns are returned.

R
owgroup Elimination
Unlike rowstore indexes, columnstore indexes have no built-in order. The data that is
inserted into a columnstore index is added in the order it is received by SQL Server.
As a result, the order of data within rowgroups is the direct result of the order it was
inserted. Equally important is that UPDATE operations will reorder a columnstore index,
removing rows from one set of rowgroups and inserting the new versions into the most
current open rowgroups.
Because compressing rowgroups is a computationally expensive process, the cost to
intrinsically maintain any form of data order would be prohibitively high. This is one of
the most important concepts when architecting a columnstore index. Since data order
is not enforced by SQL Server, it is the responsibility of the architect to determine data
order up front and ensure that both data load processes and common query patterns
maintain that agreed-upon data order.
Consider the table Dimension.Employee that contains a clustered rowstore index on
the Employee Key column. If three new employees start work and are added to the table,
the INSERT operations to add them could be represented by the T-SQL in Listing 10-4.

Listing 10-4. Query to Add Three New Employees to the Dimension.Employee

Table
INSERT INTO Dimension.Employee
([Employee Key], [WWI Employee ID], Employee, [Preferred Name],
[Is Salesperson], Photo, [Valid From], [Valid To], [Lineage Key])
VALUES
( -1, -- Clustered Index

142
Chapter 10 Segment and Rowgroup Elimination

289, N'Ebenezer Scrooge', N'Scrooge', 0, NULL, GETUTCDATE(),
'9999-12-31 23:59:59.9999999', 3),
(   213, -- Clustered Index

400, N'Captain Ahab', N'Captain', 0, NULL, GETUTCDATE(), '9999-12-31
23:59:59.9999999', 3),
(   1017, -- Clustered Index

501, N'Holden Caulfield', N'Phony', 0, NULL, GETUTCDATE(), '9999-12-31
23:59:59.9999999', 3);

There are three rows inserted into the table, with clustered index ID values of -1, 213,
and 1017. When inserted, SQL Server will place each row in the b-tree index in order
with the rest of the rows, based on those Employee Key values. As a result, the table will
remain ordered by the clustered index after the INSERT operation.
Imagine for a moment that this table did not have a rowstore index, but instead had
a clustered columnstore index. In that scenario, the three rows would be inserted at the
end of the open rowgroup(s) without any regard to the value of Employee Key. A query
that searches for a specific range of IDs will need to examine any rowgroup that contains
the IDs.
Columnstore metadata helps SQL Server locate rows based on the range of values for
each column present in each rowgroup. Consider the metadata for the Invoice Date Key
column in Fact.Sale_CCI using the query in Listing 10-5.

Listing 10-5. Query to Examine Metadata for a Single Column of a

Columnstore Index
SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       columns.name AS column_name,
       partitions.partition_number,
       column_store_segments.segment_id,
       column_store_segments.min_data_id,
       column_store_segments.max_data_id,
       column_store_segments.row_count
FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id

143
Chapter 10 Segment and Rowgroup Elimination

INNER JOIN sys.indexes

ON indexes.index_id = partitions.index_id
AND indexes.object_id = partitions.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.columns
ON tables.object_id = columns.object_id
AND column_store_segments.column_id = columns.column_id
WHERE tables.name = 'Sale_CCI'
AND columns.name = 'Invoice Date Key'
ORDER BY tables.name, columns.name, column_store_segments.segment_id;

The results are shown in Figure 10-5.

Figure 10-5. Metadata for the Invoice Date Key column of a columnstore index

Note that min_data_id and max_data_id are identical for each rowgroup. This means
that the data contained for that column is unordered. If queries commonly filtered
using Invoice Date Key, they would need to scan all rowgroups in the columnstore index
in order to appropriately filter out the requested rows. As a columnstore index grows
over time, the cost to scan all rowgroups will become burdensome. Even on a well-
compressed columnstore index, queries will become slow, and the IO required to service
an unfilterable query will be high.
STATISTICS IO provides a useful guide to the number of rowgroups read as part of a
columnstore index scan. To demonstrate this, the query in Listing 10-6 will be used.

144
Chapter 10 Segment and Rowgroup Elimination

Listing 10-6. Query to Illustrate Rowgroup Reads in a Columnstore Index Scan

SELECT
SUM([Quantity])
FROM Fact.Sale_CCI
WHERE [Invoice Date Key] >= '1/1/2016'
AND [Invoice Date Key] < '2/1/2016';

This is a classic analytic query that calculates the total sales quantity for a
given month.

Figure 10-6. STATISTICS IO for a sample unordered columnstore index scan

When a columnstore index is scanned, STATISTICS IO will include a message in

the output that indicates how many segments were read and skipped. This is a measure
of the number of rowgroups that had to be scanned to determine which rows met the
filter criteria and which could be automatically skipped. Figure 10-6 indicates that 26
segments were read and 0 were skipped. Because the underlying data is completely
unordered, SQL Server cannot use columnstore metadata to skip any rowgroups. This
represents a common real-world challenge, but one that is easy to solve.
Columnstore metadata allows rowgroups to be skipped when query filters do not
include rows present in them. This process is called rowgroup elimination and is key
to optimizing columnstore index performance. The easiest way to achieve this is by
ordering data and maintaining that order over time. Data may be ordered by one or more
columns that will represent the most common filters used in analytic queries. The most
common dimension to order OLAP data by is time. Analytics often filter, aggregate, and
visualize data using time-based units, such as hours, days, weeks, months, quarters, and
years. Figure 10-7 shows the impact of rowgroup elimination on a query that requires
only a narrow range of data based on filter criteria that honors ordered data.

145
Chapter 10 Segment and Rowgroup Elimination

Figure 10-7. Impact of rowgroup elimination on a narrow analytic query

While this hypothetical index contains six rowgroups, only a single one is required to
satisfy the filter criteria of the query. The power of rowgroup elimination is that it scales
effectively as a columnstore index grows in size. A query that requests a narrow week of
analytic data from a table with a month of data will perform similarly to that same query
against a table with 10 years of data. This is the primary feature that allows columnstore
indexes to scale effectively, even when billions of rows are present in a table.
In the example in Listing 10-6, a simple analytic query filtered on Invoice Date Key,
but the unordered columnstore index data forced a full scan of the data to determine
which rows met the filter criteria. If Invoice Date Key is the most common filter criteria
for analysis of this data, then ordering by that column would allow for effective rowgroup
elimination.

Listing 10-7. Query to Create a New Columnstore Index Ordered by Invoice

Date Key
CREATE TABLE Fact.Sale_CCI_ORDERED
( [Sale Key] [bigint] NOT NULL,
[City Key] [int] NOT NULL,

146
Chapter 10 Segment and Rowgroup Elimination

[Customer Key] [int] NOT NULL,

       [Bill To Customer Key] [int] NOT NULL,
       [Stock Item Key] [int] NOT NULL,
       [Invoice Date Key] [date] NOT NULL,
       [Delivery Date Key] [date] NULL,
       [Salesperson Key] [int] NOT NULL,
       [WWI Invoice ID] [int] NOT NULL,
       [Description] [nvarchar](100) NOT NULL,
       [Package] [nvarchar](50) NOT NULL,
       [Quantity] [int] NOT NULL,
       [Unit Price] [decimal](18, 2) NOT NULL,
       [Tax Rate] [decimal](18, 3) NOT NULL,
       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,
       [Total Dry Items] [int] NOT NULL,
       [Total Chiller Items] [int] NOT NULL,
       [Lineage Key] [int] NOT NULL);

CREATE CLUSTERED INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED

([Invoice Date Key]);

INSERT INTO Fact.Sale_CCI_ORDERED

147
Chapter 10 Segment and Rowgroup Elimination

Sale.[Salesperson Key], Sale.[WWI Invoice ID], Sale.Description,

Sale.Package, Sale.Quantity, Sale.[Unit Price], Sale.[Tax Rate],
Sale.[Total Excluding Tax], Sale.[Tax Amount], Sale.Profit,
Sale.[Total Including Tax], Sale.[Total Dry Items],
Sale.[Total Chiller Items], Sale.[Lineage Key]
FROM fact.Sale
CROSS JOIN
Dimension.City
WHERE City.[City Key] >= 1 AND City.[City Key] <= 110;

-- Create a clustered columnstore index on the table, removing the existing

clustered rowstore index.
CREATE CLUSTERED COLUMNSTORE INDEX CCI_fact_Sale_CCI_ORDERED ON
Fact.Sale_CCI_ORDERED WITH (MAXDOP = 1, DROP_EXISTING = ON);
GO

Note that the data in the table created in Listing 10-7 is identical to the data
demonstrated earlier in this chapter, but has been subject to a clustered rowstore index
prior to being given a columnstore index. This additional step ensures that the initial
data set is ordered by Invoice Date Key. MAXDOP is intentionally set to 1 to avoid
parallelism as parallel threads may risk inserting data into the columnstore index in
multiple ordered streams rather than a single ordered stream.
Going forward, new data would be regularly inserted into this table via standard
data load processes. Assuming the new data contains the most recent values for
Invoice Date Key, then the columnstore index will remain ordered in the future as new
data is added to it.
To test the impact of data order on Fact.Sale_CCI_ORDERED, the query from
Listing 10-6 will be executed against it, with the output tab displayed in Figure 10-8.

Figure 10-8. STATISTICS IO for a sample filtered columnstore index scan

148
Chapter 10 Segment and Rowgroup Elimination

Instead of reading every rowgroup in the columnstore index, SQL Server only
needed to read two of them, with the remainder being skipped. Skipped segments in
STATISTICS IO indicate that rowgroup elimination is being successfully implemented.
The metadata query in Listing 10-5 can also be rerun against this ordered table to
illustrate how data order impacts columnstore metadata, with the results being shown in
Figure 10-9.

Figure 10-9. Metadata for the Invoice Date Key column of an ordered
columnstore index

The values of min_data_id and max_data_id for each rowgroup show a drastic
change from that of an unordered columnstore index. Instead of the values being the
same for each rowgroup, they progress from low values to high values as data progresses
from the first rowgroup to the latter ones. To put this in perspective, if a hypothetical
query required data for a data ID of 735270, it would only need to read the rowgroups
associated with segment_id = 7 from this list of segments. Since the metadata indicates
that the remaining segments do not contain this value, they (and their associated
rowgroups) can automatically be skipped.
Ordering data by a key column is an easy way to enable rowgroup elimination,
thereby reducing query resource consumption and improving the speed of any queries
that can make use of that data order. Effective data order doesn’t just reduce reads, but
it can also save on storage space by improving the compression ratios of the underlying
data. Listing 10-8 contains a script that will retrieve the data space used by the two
columnstore indexes featured in this chapter.

149
Chapter 10 Segment and Rowgroup Elimination

Listing 10-8. Query to Retrieve Data Space Used for Two Columnstore Indexes

CREATE TABLE #storage_data

(      table_name VARCHAR(MAX),
       rows_used BIGINT,
       reserved VARCHAR(50),
       data VARCHAR(50),
       index_size VARCHAR(50),
       unused VARCHAR(50));

INSERT INTO #storage_data

(table_name, rows_used, reserved, data, index_size, unused)
EXEC sp_MSforeachtable "EXEC sp_spaceused '?'";

UPDATE #storage_data
       SET reserved = LEFT(reserved, LEN(reserved) - 3),
              data = LEFT(data, LEN(data) - 3),
              index_size = LEFT(index_size, LEN(index_size) - 3),
              unused = LEFT(unused, LEN(unused) - 3);
SELECT
       table_name,
       rows_used,
       reserved / 1024 AS data_space_reserved_mb,
       data / 1024 AS data_space_used_mb,
       index_size / 1024 AS index_size_mb,
       unused AS free_space_kb,
       CAST(CAST(data AS DECIMAL(24,2)) / CAST(rows_used AS DECIMAL(24,2))
AS DECIMAL(24,4)) AS kb_per_row
FROM #storage_data
WHERE rows_used > 0
AND table_name IN ('Sale_CCI', 'Sale_CCI_ORDERED')
ORDER BY CAST(reserved AS INT) DESC;

DROP TABLE #storage_data;

The results are shown in Figure 10-10.

150
Chapter 10 Segment and Rowgroup Elimination

Figure 10-10. Space used by an ordered and unordered columnstore index

The data space used for each table is dramatically different, with the ordered
table consuming less than 10% of the space that the unordered table uses. This is an
exceptionally dramatic example of how ordered data can be stored more efficiently
than unordered data. Ordered data saves space because, typically, data is more similar
to other data captured within a short timeframe of it than when compared to data that
was collected years apart. Within any application, usage patterns change over time as
new features are released, old features are retired, and user behavior changes. Because
of this, data samples will look more and more different as time passes between them.
These similarities translate into compression algorithms being able to take advantage of
a data set with less distinct values for common dimensions. This also reduces dictionary
size, which also helps to prevent dictionaries from filling up and forcing the creation of
undersized rowgroups.
Real-world data may not compress as impressively as the sample here, but expect
nontrivial savings that will have a positive impact on data load processes and on
analytics speeds. It is important to remember that saving storage space also saves
memory as data remains compressed until needed by an application. Therefore, if an
ordered data set were to decrease in size by 25%, that would result in 25% less memory
being consumed in the buffer pool by columnstore index pages. Furthermore, other
common measures of server performance such as page life expectancy and latching
would improve as smaller objects can be retrieved more quickly and will impact other
data in memory less than larger objects.
An ordered columnstore index also improves UPDATE and DELETE speeds by
allowing those operations to target less pages. For example, consider the queries on the
ordered and unordered sales tables shown in Listing 10-9.

151
Chapter 10 Segment and Rowgroup Elimination

Listing 10-9. Query to Display Sample Row Counts for Two Columnstore
Indexed Tables
SELECT COUNT(*) AS row_count FROM Fact.Sale_CCI WHERE [Invoice Date Key] =
'1/1/2015';
SELECT COUNT(*) AS row_count FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date
Key] = '1/1/2015';

The results show that in each table, the count of rows affected is identical, as seen in
Figure 10-11.

Figure 10-11. Row counts for a sample query against ordered and unordered
columnstore indexes

For each table, 33,110 rows would be affected by an update using the same filter.
Listing 10-10 provides a simple UPDATE statement against each table.

Listing 10-10. Query to Update 33,110 Rows in Two Columnstore Indexed Tables

UPDATE Sale_CCI
       SET [Total Dry Items] = [Total Dry Items] - 1,
              [Total Chiller Items] = [Total Chiller Items] + 1
FROM Fact.Sale_CCI_ORDERED -- Unordered
WHERE [Invoice Date Key] = '1/1/2015';
UPDATE Sale_CCI
       SET [Total Dry Items] = [Total Dry Items] - 1,
              [Total Chiller Items] = [Total Chiller Items] + 1
FROM Fact.Sale_CCI_ORDERED -- Ordered
WHERE [Invoice Date Key] = '1/1/2015';

The results in Figure 10-12 show the resulting IO and rowgroup usage for each
UPDATE operation.

152
Chapter 10 Segment and Rowgroup Elimination

Figure 10-12. STATISTICS IO output for updates against an ordered and

unordered columnstore index

Note the dramatic difference in IO, as well as segment reads for each operation.
Because an UPDATE consists of both a DELETE and an INSERT, SQL Server had to
perform the following tasks to complete each update:
1. Locate all rows matching the filter criteria.
2. Read all columns for all rows matching the filter criteria.
3. Mark these rows as deleted in the delete bitmap.
4. Insert new versions of these rows into the delta store.
In order to insert new versions of the updated rows, SQL Server needs to read the
existing rows in their entirety, which is not a trivial operation. Once read, those rows
are marked as deleted and the new versions are inserted into the delta store. This is an
expensive process, but ordered data allows for far fewer rowgroups to be read, thereby
reducing the work needed to set up the necessary data for the insert into the delta store.
DELETE operations are improved similarly, but require far less work as they
simply need to
1. Locate all rows matching the filter criteria.
2. Mark these rows as deleted in the delete bitmap.

For both cases, an ordered columnstore index will immensely improve UPDATE
and DELETE performance when the filter criteria honors the order used in the table.
As an added bonus, an ordered columnstore index will allow for DELETE and UPDATE
operations that cause less fragmentation. Instead of flagging rows as deleted in most (or
all) rowgroups, the deletes can be isolated to a smaller number of rowgroups.
153
Chapter 10 Segment and Rowgroup Elimination

Combining Segment and Rowgroup Elimination

Segment elimination allows analytics that use less columns to automatically require less
segments to be read, thereby reducing IO. Rowgroup elimination allows an ordered data
set to facilitate analytic queries to use columnstore metadata to remove vast swaths of
rowgroups from the result set, reducing IO significantly.
Optimal analytic workloads combine segment and rowgroup elimination to slice a
table both vertically (rowgroup elimination) and horizontally (segment elimination).
Consider the columnstore index introduced previously in Figure 10-1. If an analytic
query were designed to only query for columns A and B and to also filter for a narrow
date range, the resulting rowgroup reads would look similar to the results in Figure 10-13.

Figure 10-13. Impact of segment and rowgroup elimination on a

columnstore index

154
Chapter 10 Segment and Rowgroup Elimination

Only querying for columns A and B allows columns C–H to be automatically skipped
and all segments for those columns eliminated (36 segments in total). Querying for a
narrow date range that only requires rows in rowgroup 4 allows rowgroups 1–3 and 5–6
to also be automatically skipped, eliminating another 10 segments. The result is a query
that required only 2 out of the possible 48 segments in the table!
Combining segment and rowgroup elimination allows columnstore index queries to
scale effectively, even as data grows larger or as more columns are added to the table.

155
CHAPTER 11

Partitioning
As an analytic table grows in size, it becomes apparent that newer data is read far more
often than older data. While columnstore metadata and rowgroup elimination provide
the ability to quickly filter out large amounts of columnstore index data, managing a
table with millions or billions of rows can become cumbersome.
Equally important is the fact that older data tends to not change often. For a typical
fact table, data is added onto the end of it in the order it is created, whereas older data
remains untouched. If older data is modified, it is usually the result of software releases,
upgrades, or other processes that fall squarely into the world of the data architect
(that’s us!) to manage.
Table partitioning is a natural fit for a clustered columnstore index, especially if
it is large. Partitioning allows data to be split into multiple filegroups within a single
database. These filegroups can be stored in different data files that reside in whatever
locations are ideal for the data contained within them. The beauty of partitioning is
that the table is logically unchanged, with its physical structure being influenced by the
details of how it partitioned. This means that application code does not need to change
in order to benefit from it. There are many benefits to partitioning, each of which is
described in this chapter.

Maintain Hot/Warm/Cold Data

If older data is used far less often, then it can be stored in separate files that reside on
slower storage. For example, reporting data from 10 years ago that is maintained for
posterity but rarely used can be placed on inexpensive NAS storage, whereas new data
can go on speedy SSD or flash storage.
The details will vary depending on the organization, but Figure 11-1 provides a
simple representation of how an analytic table could be tiered to take advantage of
different service levels.

157
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_11
Chapter 11 Partitioning

Figure 11-1. How partitioning can influence storage speed and cost

Partitioning allows storage to be tiered based on expected SLAs (service-level

agreements). For partitions containing new data that is expected to be highly available
and have low latency, fast and expensive storage can be used. For partitions containing
older data that is rarely accessed, slower and cheaper storage can be used. The
details are up to an organization, but the ability to efficiently divide up a large table
automatically into different storage tiers is exceptionally useful and can save significant
money in the long run.

Faster Data Movement/Migration

Having data divided up physically means that it can be copied and moved with ease.
For example, consider a columnstore index with 50 billion rows that consumes 1TB of
storage. If there was a need to migrate this table from its current server to a new server
with minimal downtime, how would it be done? The simplest solutions involve database
backups, copying data files, or using ETL to move the data slowly from one server to the
other. All of these options work, but would be time-consuming and would involve the
need to incur some downtime or make the table read-only once the data movement
process starts.
Partitioning allows older/unchanging data to reside in separate data files, which in
turn means that those files can be freely backed up/copied to the new server ahead of
time. Assuming nothing changes within them, that process can occur days, weeks, or

158
Chapter 11 Partitioning

months prior to the migration. ETL or similar processes could be used on the day of the
migration to catch up the target database with new data prior to permanently cutting
over from the old data source to the new one.
Figure 11-2 illustrates the difference between migrating a large/monolithic table vs. a
partitioned one.

Figure 11-2. Migration of a partitioned table vs. a nonpartitioned table

Partitioning opens up the ability to subdivide the migration logically knowing that
the physical storage of the table will facilitate the ability to copy/move each file one by
one when needed. Instead of having to move a terabyte of data all at once or being forced
to write ETL against the entire table, the table can be subdivided into smaller pieces,
each of which is smaller and easier to move.

P
artition Elimination
The logical definitions for each partition are not solely used for storage purposes. They
also assist the query optimizer and allow it to automatically skip data when the filter in
a query aligns with the partition function. For a query against a columnstore index, its
filter can allow the metadata for rowgroups outside of the target partition to be ignored.

159
Chapter 11 Partitioning

For example, if a columnstore index contains data ranging from 2010 through 2021
and is partitioned by year (with a single partition per year), then a query requesting rows
from January 2021 would be able to automatically skip all rowgroups in partitions prior
to 2021. Figure 11-3 shows how partition elimination can reduce reads and further speed
up queries against columnstore indexes.

Figure 11-3. Reading metadata in a partitioned columnstore index

Partition functions are evaluated prior to columnstore metadata; therefore, data

in irrelevant partitions is ignored when a query is executed. While columnstore
metadata may be relatively lightweight, being able to skip the metadata for thousands
of rowgroups can further improve analytic performance while reducing IO and memory
consumption.

D
atabase Maintenance
Some database maintenance tasks, such as backups and index maintenance, can be
executed on a partition-by-partition basis. Since the data in older partitions rarely
changes, it is less likely to require maintenance as often as newer data. Therefore,

160
Chapter 11 Partitioning

maintenance can be focused on the specific partitions that need it most. This also means
that maintenance speeds can be greatly improved by no longer needing to operate on
the entire table at one time.
In addition, how data is managed can vary partition to partition. Some ways in which
data can be handled differently depending on the partition include

• Compression can be customized by partition. In a columnstore

index, archive compression can be applied to older and less used
partitions, whereas standard columnstore compression can be the
default for newer data.

• Partitions can be truncated individually, allowing data to be

removed from a portion of a table quickly, without having to use a
DELETE statement.

• Partition switching allows data to be swapped in and out of a

partitioned table quickly and efficiently. This can greatly speed up
migrations, data archival, and other processes where a large volume
of data needs to be moved at one time.

P
artitioning in Action
To visualize how partitioning works, a new version of the test table Fact.Sale_CCI will be
created. This version will be ordered by Invoice Date Key and also partitioned by year.
Each partition needs to target a filegroup, which in turn will contain a data file. For this
demonstration, a new filegroup and file will be created for each year represented in
the table.
Listing 11-1 shows how new filegroups can be created that will be used to store data
for the partitioned table.

Listing 11-1. Script That Creates a New Filegroup for Each Year of Data in
the Table
ALTER DATABASE WideWorldImportersDW ADD FILEGROUP
WideWorldImportersDW_2013_fg;
ALTER DATABASE WideWorldImportersDW ADD FILEGROUP
WideWorldImportersDW_2014_fg;

161
Chapter 11 Partitioning

ALTER DATABASE WideWorldImportersDW ADD FILEGROUP

WideWorldImportersDW_2015_fg;
ALTER DATABASE WideWorldImportersDW ADD FILEGROUP
WideWorldImportersDW_2016_fg;
ALTER DATABASE WideWorldImportersDW ADD FILEGROUP
WideWorldImportersDW_2017_fg;

Once executed, the presence of the new database filegroups can be confirmed by
checking the Filegroups menu within the database’s properties, as seen in Figure 11-4.

Figure 11-4. New filegroups that will be used to store partitioned data

While the five new filegroups are present in the database, they contain no files. The
next step in this process is to add files to these filegroups (one file each). Listing 11-2
contains the code needed to add these files.

Listing 11-2. Script to Add Files to Each New Filegroup

ALTER DATABASE WideWorldImportersDW ADD FILE

       (NAME = WideWorldImportersDW_2013_data, FILENAME =
'C:\SQLData\WideWorldImportersDW_2013_data.ndf',
        SIZE = 200MB, MAXSIZE = UNLIMITED, FILEGROWTH = 1GB)
TO FILEGROUP WideWorldImportersDW_2013_fg;
ALTER DATABASE WideWorldImportersDW ADD FILE
       (NAME = WideWorldImportersDW_2014_data, FILENAME =
'C:\SQLData\WideWorldImportersDW_2014_data.ndf',
        SIZE = 200MB, MAXSIZE = UNLIMITED, FILEGROWTH = 1GB)
TO FILEGROUP WideWorldImportersDW_2014_fg;

162
Chapter 11 Partitioning

ALTER DATABASE WideWorldImportersDW ADD FILE

(NAME = WideWorldImportersDW_2015_data, FILENAME =
'C:\SQLData\WideWorldImportersDW_2015_data.ndf',

SIZE = 200MB, MAXSIZE = UNLIMITED, FILEGROWTH = 1GB)
TO FILEGROUP WideWorldImportersDW_2015_fg;
ALTER DATABASE WideWorldImportersDW ADD FILE

(NAME = WideWorldImportersDW_2016_data, FILENAME =
'C:\SQLData\WideWorldImportersDW_2016_data.ndf',

SIZE = 200MB, MAXSIZE = UNLIMITED, FILEGROWTH = 1GB)
TO FILEGROUP WideWorldImportersDW_2016_fg;
ALTER DATABASE WideWorldImportersDW ADD FILE

(NAME = WideWorldImportersDW_2017_data, FILENAME =
'C:\SQLData\WideWorldImportersDW_2017_data.ndf',

SIZE = 200MB, MAXSIZE = UNLIMITED, FILEGROWTH = 1GB)
TO FILEGROUP WideWorldImportersDW_2017_fg;

For this demonstration, each file has the same size and growth settings, but for a
real-world table, these numbers will vary. By inspecting the size of the data that is to
be partitioned, the amount of space needed for each year should be relatively easy to
calculate. Note that once a table’s data is migrated to a partitioned table, the data files
that used to contain its data will now have additional free space that can be reclaimed, if
needed. Figure 11-5 shows the new files listed under the Files menu within the database
properties.

Figure 11-5. New database files that will be used to store partitioned data

163
Chapter 11 Partitioning

New database files can be placed on any storage available to the server. This is
where the table’s data can be customized to meet whatever SLAs it is subject to. For
this example, if data from prior to 2016 is rarely accessed, it could be placed into files
on slower storage. Similarly, more recent data could be maintained on faster storage to
support frequent analytics.
The next step in configuring partitioning is to determine how to slice up data from
the table into the newly created files. This is achieved using a partition function. When
working with a columnstore index, ensure that the column that the table is ordered by
is also the same data type used for the partition function. This convention is the same
with clustered rowstore indexes as well. If the data type for the partition function does
not match the target column to order by in the table, an error will be thrown when the
table is created. For this demonstration, the partition function will split up data using the
DATE data type, which corresponds to the Invoice Date Key column that Fact.Sale_CCI is
ordered by, as seen in Listing 11-3.

Listing 11-3. Creation of a Partition Function That Organizes Data by Date

CREATE PARTITION FUNCTION fact_Sale_CCI_years_function (DATE)

AS RANGE RIGHT FOR VALUES
('1/1/2014', '1/1/2015', '1/1/2016', '1/1/2017');

RANGE RIGHT specifies that the boundaries created will be defined using the dates
provided as starting points. The result is that data will be divided into five buckets,
like this:

1. All dates < 1/1/2014

2. Dates >= 1/1/2014 and < 1/1/2015

3. Dates >= 1/1/2015 and < 1/1/2016

4. Dates >= 1/1/2016 and < 1/1/2017

5. Dates >= 1/1/2017

If used, RANGE LEFT would result in date ranges where the inequality operators
are adjusted so that boundaries are checked with greater than and less than or equal to,
rather than what was presented earlier. If unsure of which to use, consider implementing
RANGE RIGHT for time-based dimensions as it is typically a more natural division that

164
Chapter 11 Partitioning

cleanly divides up units of months, quarters, years, etc. The following list shows how the
boundaries would be defined in the partition function in Listing 11-3 if RANGE LEFT
were used:

1. All dates <= 1/1/2014

2. Dates > 1/1/2014 and <= 1/1/2015

3. Dates > 1/1/2015 and <= 1/1/2016

4. Dates > 1/1/2016 and <= 1/1/2017

5. Dates > 1/1/2017

This boundary configuration is somewhat counterintuitive for dates, but might be

relevant for other data types.
The final step before partitioning can be implemented on a table is to create a
partition scheme. A partition scheme is used to map the data boundaries defined in the
partition function onto the filegroups defined earlier. This is where it can be decided
which date ranges correspond to which physical database files. Listing 11-4 provides a
partition scheme that assigns each date range defined by the partition function to one of
the five newly created filegroups.

Listing 11-4. Creation of a Partition Scheme to Assign Dates to Database

Filegroups
CREATE PARTITION SCHEME fact_Sale_CCI_years_scheme
AS PARTITION fact_Sale_CCI_years_function
TO (WideWorldImportersDW_2013_fg, WideWorldImportersDW_2014_fg,
WideWorldImportersDW_2015_fg, WideWorldImportersDW_2016_fg,
WideWorldImportersDW_2017_fg);

Note that the partition function will always specify one less boundary than there
are filegroups. In this example, the partition function provides four dates that form date
boundaries that define five distinct date ranges which are subsequently mapped onto
the partition scheme and assigned filegroups. If the number of boundaries provided by
the partition function is not one less than the partition scheme, an error will be returned,
similar to what is seen in Figure 11-6.

165
Chapter 11 Partitioning

Figure 11-6. Error received if partition scheme contains too few filegroup entries

The error is verbose enough to remind the user that the function generates too
many or too few partitions than are afforded by the partition scheme. By writing out the
time periods desired for the target table, the task of generating a partition function and
partition scheme become far simpler.
With database files and filegroups defined, as well as a partition function and
scheme, the final step to implementing partitioning is to create a table using the newly
created partition scheme. Listing 11-5 creates a new version of Fact.Sale and partitions it
using the Invoice Date Key column on fact_Sale_CCI_years_scheme.

Listing 11-5. Creation of a New Table That Is Partitioned by Year on a

Date Column
CREATE TABLE Fact.Sale_CCI_PARTITIONED
(      [Sale Key] [bigint] NOT NULL,
       [City Key] [int] NOT NULL,
       [Customer Key] [int] NOT NULL,
       [Bill To Customer Key] [int] NOT NULL,
       [Stock Item Key] [int] NOT NULL,
       [Invoice Date Key] [date] NOT NULL,
       [Delivery Date Key] [date] NULL,
       [Salesperson Key] [int] NOT NULL,
       [WWI Invoice ID] [int] NOT NULL,
       [Description] [nvarchar](100) NOT NULL,
       [Package] [nvarchar](50) NOT NULL,
       [Quantity] [int] NOT NULL,
       [Unit Price] [decimal](18, 2) NOT NULL,
       [Tax Rate] [decimal](18, 3) NOT NULL,
       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,

166
Chapter 11 Partitioning

[Total Dry Items] [int] NOT NULL,

       [Total Chiller Items] [int] NOT NULL,
       [Lineage Key] [int] NOT NULL)
       ON fact_Sale_CCI_years_scheme ([Invoice Date Key]);

The only difference in the definition of a partitioned table vs. a nonpartitioned

table is the last line, where an ON clause defines the partition scheme to use and which
column will be evaluated to determine which filegroups the data will be stored in. The
data type of the partition column and partition function must match or an error will be
returned, similar to what is shown in Figure 11-7.

Figure 11-7. Error received if the data type of the partition column and partition
function do not match

Data types between the partition function and column must match exactly. DATE
and DATETIME are not compatible, nor are other data types that may seem similar. SQL
Server does not automatically convert between data types when evaluating partition
functions and will instead throw an error when the table is created.
Once created, Fact.Sale_CCI_PARTITIONED is managed in the same fashion as
Fact.Sale_CCI_ORDERED was:

1. A clustered rowstore index is created, ordered by Invoice Date Key.

2. Data is inserted into the table.

3. A clustered columnstore index is created, replacing the

rowstore index.

The code in Listing 11-6 walks through each of these steps.

167
Chapter 11 Partitioning

Listing 11-6. Populating a Partitioned Table with an Ordered Data Set

CREATE CLUSTERED INDEX CCI_fact_Sale_CCI_PARTITIONED ON

Fact.Sale_CCI_PARTITIONED ([Invoice Date Key]);

INSERT INTO Fact.Sale_CCI_PARTITIONED

       ([Sale Key], [City Key], [Customer Key], [Bill To Customer Key],
[Stock Item Key], [Invoice Date Key], [Delivery Date Key],
        [Salesperson Key], [WWI Invoice ID], Description, Package,
Quantity, [Unit Price], [Tax Rate],
        [Total Excluding Tax], [Tax Amount], Profit, [Total Including Tax],
[Total Dry Items],
        [Total Chiller Items], [Lineage Key])
SELECT
       Sale.[Sale Key], Sale.[City Key], Sale.[Customer Key], Sale.[Bill To
Customer Key], Sale.[Stock Item Key], Sale.[Invoice Date Key],
Sale.[Delivery Date Key],
       Sale.[Salesperson Key], Sale.[WWI Invoice ID], Sale.Description,
Sale.Package, Sale.Quantity, Sale.[Unit Price], Sale.[Tax Rate],
       Sale.[Total Excluding Tax], Sale.[Tax Amount], Sale.Profit,
Sale.[Total Including Tax], Sale.[Total Dry Items],
       Sale.[Total Chiller Items], Sale.[Lineage Key]
FROM Fact.Sale
CROSS JOIN
Dimension.City
WHERE City.[City Key] >= 1 AND City.[City Key] <= 110;

CREATE CLUSTERED COLUMNSTORE INDEX CCI_Fact_Sale_CCI_PARTITIONED ON

Fact.Sale_CCI_PARTITIONED WITH (MAXDOP = 1, DROP_EXISTING = ON);

When complete, two tables will exist for demonstration purposes that are identical,
except that one is partitioned and the other is not. Both are ordered by Invoice Date Key
and will benefit from rowgroup elimination whenever filtered by that column. Listing 11-7
shows two queries that aggregate Quantity for a single month against each table.

168
Chapter 11 Partitioning

Listing 11-7. Narrow Analytic Queries Executed Against Nonpartitioned and

Partitioned Columnstore Indexes
SELECT
SUM([Quantity])
FROM Fact.Sale_CCI_ORDERED
WHERE [Invoice Date Key] >= '1/1/2016'
AND [Invoice Date Key] < '2/1/2016';

SELECT
SUM([Quantity])
FROM Fact.Sale_CCI_PARTITIONED
WHERE [Invoice Date Key] >= '1/1/2016'
AND [Invoice Date Key] < '2/1/2016';

While the queries are identical aside from the table name, the STATISTICS IO output
illustrates the differences in execution between each, as seen in Figure 11-8.

Figure 11-8. STATISTICS IO for a nonpartitioned columnstore index vs. a

partitioned columnstore index

The output of STATISTICS IO shows multiple ways in which query execution varied
for each table. The most significant difference is in the reported segment reads. The
nonpartitioned table read 2 segments while skipping 22, whereas the partitioned table
read 1 segment while skipping 3. This IO reflects the original query that calculates a sum
using only rows with an Invoice Date Key within January 2016.
For the table that is ordered and not partitioned, metadata needs to be reviewed
from the entire table prior to using rowgroup elimination to skip segments. In
the partitioned table, though, rowgroups that do not contain data from 2016 are

169
Chapter 11 Partitioning

automatically skipped. Since partition functions are evaluated prior to evaluating

columnstore metadata, unrelated rowgroups are never read, including their metadata.
This represents partition elimination in action. While partitioning is not intended to be a
query optimization solution, it will have a positive impact on columnstore index IO and
query speeds, especially on larger tables.
Partitioning also impacts the query optimizer, which will evaluate a query across
less rowgroups and rows, which can simplify execution plans. Figure 11-9 shows the
execution plans for the analytic queries demonstrated in Listing 11-7.

Figure 11-9. Query execution plans for a nonpartitioned columnstore index vs. a
partitioned columnstore index

The execution plan for the nonpartitioned table is more complex as SQL Server
determines that parallelism may help in processing the large number of rows. The
execution plan for the partitioned table is simpler, as the optimizer chooses to not use
parallelism. Note that the use of a hash match is to support the aggregate pushdown
for the SUM operator into the columnstore index scan step. Also interesting to note is
that the count of rows read in the execution plan is lower on a partitioned table (vs. a
nonpartitioned table) when less partitions need to be read.
While the output of each query is identical and each execution plan produces the
same result set, the ability to forgo parallelism saves computing resources, as is implied
by the greatly reduced query cost for the partitioned table.
In addition to partition elimination, index maintenance can be adjusted to skip
older, less updated partitions. For example, if a rebuild was deemed necessary for this
columnstore index to clean it up after some recent software releases, the nonpartitioned
table would need to be rebuilt en masse, as seen in Listing 11-8.
170
Chapter 11 Partitioning

Listing 11-8. Query to Rebuild a Columnstore Index

ALTER INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED REBUILD;

The rebuild operation takes 61 seconds to complete. For a larger columnstore index,
it could be significantly longer. If the portion of data impacted by the software release
is limited to newer data only, it is very likely that only the most current partition needs
to be rebuilt. Listing 11-9 shows how an index rebuild can be executed against a single
partition.

Listing 11-9. Query to Rebuild the Current Partition in a Columnstore Index

ALTER INDEX CCI_fact_Sale_CCI_PARTITIONED ON Fact.Sale_CCI_PARTITIONED

REBUILD PARTITION = 5;

This rebuild only takes 1 second. This is because the most recent partition does not
contain a full year of data. To provide a fairer assessment of rebuild times, Listing 11-10
shows a rebuild operation against a partition that contains a full year of data.

Listing 11-10. Query to Rebuild a Full Partition in a Columnstore Index

ALTER INDEX CCI_fact_Sale_CCI_PARTITIONED ON Fact.Sale_CCI_PARTITIONED

REBUILD PARTITION = 4;

This rebuild operation requires 11 seconds to complete. Each partition rebuild is

significantly faster than being forced to rebuild the entire index. This can greatly reduce
the performance impact of index maintenance, when it is needed.
Another unique feature that can be demonstrated is the use of the SWITCH option
to move the contents of a partition from one table to another using a single DDL
operation. Because the operation is a DDL metadata operation and not a fully logged
data movement process, it will be significantly faster. This opens up a variety of options
for speedy data load and maintenance processes that otherwise would be prohibitively
expensive.
For example, if it were determined that all data in Fact.Sale_CCI_PARTITIONED had
an error starting on 1/1/2016, correcting it would be a challenge if left in place. Updating
a columnstore index is a slow and expensive process that results in heavy fragmentation
and is therefore worth avoiding if possible.

171
Chapter 11 Partitioning

An alternative to this would be to switch out the partition, modify the data, and then
reinsert the data back into the table. This allows the data that is to be modified to be
isolated prior to making any changes. The script in Listing 11-11 creates a new staging
table with the same schema as Fact.Sale_CCI_PARTITIONED that will be used as the
target for a partition switch.

Listing 11-11. Script to Create a Staging Table for Use in a Partition Switching
Operation
CREATE TABLE Fact.Sale_CCI_STAGING
(      [Sale Key] [bigint] NOT NULL,
       [City Key] [int] NOT NULL,
       [Customer Key] [int] NOT NULL,
       [Bill To Customer Key] [int] NOT NULL,
       [Stock Item Key] [int] NOT NULL,
       [Invoice Date Key] [date] NOT NULL,
       [Delivery Date Key] [date] NULL,
       [Salesperson Key] [int] NOT NULL,
       [WWI Invoice ID] [int] NOT NULL,
       [Description] [nvarchar](100) NOT NULL,
       [Package] [nvarchar](50) NOT NULL,
       [Quantity] [int] NOT NULL,
       [Unit Price] [decimal](18, 2) NOT NULL,
       [Tax Rate] [decimal](18, 3) NOT NULL,
       [Total Excluding Tax] [decimal](18, 2) NOT NULL,
       [Tax Amount] [decimal](18, 2) NOT NULL,
       [Profit] [decimal](18, 2) NOT NULL,
       [Total Including Tax] [decimal](18, 2) NOT NULL,
       [Total Dry Items] [int] NOT NULL,
       [Total Chiller Items] [int] NOT NULL,
              [Lineage Key] [int] NOT NULL)
ON WideWorldImportersDW_2016_fg;
CREATE CLUSTERED COLUMNSTORE INDEX CCI_fact_Sale_CCI_STAGING ON
Fact.Sale_CCI_STAGING;

172
Chapter 11 Partitioning

Note that the staging table is created on the same filegroup as the source data that
is to be switched. If desired, the staging table could be created using the same partition
scheme as Fact.Sale_CCI_PARTITIONED, which would allow the partition number to be
specified in the partition switch, rather than having to explicitly provide a filegroup in
the table create statement. The syntax is arbitrary and can be left to the convenience of
the operator as to which is easier to implement.
Once the staging table is created, a partition on the 2016 filegroup can be switched
from the source table into the staging table using the script in Listing 11-12.

Listing 11-12. Script to Switch a Partition from a Partitioned Columnstore Index

into a Staging Table
ALTER TABLE Fact.Sale_CCI_PARTITIONED SWITCH PARTITION 4 TO
Fact.Sale_CCI_STAGING;

The partition switch executes in milliseconds as it is a minimally logged DDL

operation. The results can be measured by validating the row count in the staging table,
as shown in Listing 11-13.

Listing 11-13. Validating the Row Count in the Staging Table After the Partition
Switch Operation Completes
SELECT COUNT(*) FROM Fact.Sale_CCI_STAGING;

The results of the count operation are shown in Figure 11-10.

Figure 11-10. Count of rows in the newly populated staging table

The result of the count query shows that over 3 million rows were switched from
Fact.Sale_CCI_PARTITIONED to the staging table. From here, data can be modified in
the staging table as needed to resolve whatever the identified issues are. Once the data is
cleaned up, it can be inserted back into the original columnstore index, ensuring that the
resulting data is clean with no fragmentation. Listing 11-14 shows a sample process for
how this data modification could occur.

173
Chapter 11 Partitioning

Listing 11-14. Example of Data Modification Using Staging Data. Data Is Moved
Back into the Parent Table Using an INSERT Operation
UPDATE Sale_CCI_STAGING
SET [Total Dry Items] = [Total Dry Items] + 1
FROM Fact.Sale_CCI_STAGING;

INSERT INTO Fact.Sale_CCI_PARTITIONED

       ([Sale Key], [City Key], [Customer Key], [Bill To Customer Key],
[Stock Item Key], [Invoice Date Key], [Delivery Date Key],
[Salesperson Key],
     [WWI Invoice ID], Description, Package, Quantity, [Unit Price],
[Tax Rate], [Total Excluding Tax], [Tax Amount], Profit, [Total
Including Tax],
     [Total Dry Items], [Total Chiller Items], [Lineage Key])
SELECT
       [Sale Key], [City Key], [Customer Key], [Bill To Customer Key],
[Stock Item Key], [Invoice Date Key], [Delivery Date Key],
[Salesperson Key],
     [WWI Invoice ID], Description, Package, Quantity, [Unit Price],
[Tax Rate], [Total Excluding Tax], [Tax Amount], Profit, [Total
Including Tax],
     [Total Dry Items], [Total Chiller Items], [Lineage Key]
FROM Fact.Sale_CCI_STAGING;

DROP TABLE Fact.Sale_CCI_STAGING;

The update is an expensive operation, but by isolating it to a staging table, the

contention of the operation will not impact the larger columnstore index. Once
the update is complete, the data can be inserted back into the partitioned table and the
staging table dropped.
Alternatively, the INSERT operation can be replaced with another partition switch.
The code in Listing 11-15 shows an alternative to the code in Listing 11-14.

174
Chapter 11 Partitioning

Listing 11-15. Example of Data Modification Using Staging Data. Data Is Moved
Back into the Parent Table Using Partition Switching
UPDATE Sale_CCI_STAGING
SET [Total Dry Items] = [Total Dry Items] + 1
FROM Fact.Sale_CCI_STAGING;

ALTER TABLE Fact.Sale_CCI_STAGING SWITCH PARTITION 4 TO

Fact.Sale_CCI_PARTITIONED;

DROP TABLE Fact.Sale_CCI_STAGING;

This alternative code will be significantly faster as the partition switch is a speedy,
minimally logged DDL operation, whereas the large INSERT in Listing 11-14 needs to
incur the cost of writing all of this data back to Fact.Sale_CCI_Partitioned.
Partition switching is a versatile tool that can be used in a wide variety of data
modification, archival, and deletion scenarios. Its use will vary between applications,
but provides a fast way to shift data in and out of tables without the fragmentation and
expense of a large write operation.

P
artitioning Guidelines
While partitioning may sound like a win for any columnstore index, it should not be
automatically implemented without research and foresight. Partitioning works well
in larger tables, but does not provide value to small columnstore indexes. Therefore,
capacity planning should be a step in the decision-making process to determine whether
partitioning is a good fit or not for any given table.

Partition and Rowgroup Sizing

Because partitioning is applied to a table before rowgroups are created, partition
boundaries will force rowgroups to be cut off. For example, if a nonpartitioned
columnstore index contains 5,000,000 rows, those rows will likely reside in 5 rowgroups.
If a new version of that table were created and partitioned on a date column that evenly
spanned 10 years, then the result would be 500,000 rows per partition.

175
Chapter 11 Partitioning

Rowgroups cannot span partitions, though. The new partitioned table would contain
ten partitions, each of which consists of a single rowgroup. The result is a table with ten
rowgroups instead of five.
Generally speaking, partitioning is not appropriate for a columnstore index unless
the table contains at least tens or hundreds of millions of rows. Equally important,
partitions need to be large enough to facilitate the use of filled-up rowgroups. Therefore,
the partition function and partition scheme need to ensure that each partition contains
at least 220 (1,048,576) rows each. Undersized partitions will result in undersized
rowgroups, which will lead to fragmentation over time.
The larger a columnstore index, the more it will benefit from partitioning. Aligning
partition boundaries to organizational needs can also help in determining how to
implement partitioning on a columnstore index. If an organization archives a billion
rows of data each quarter, then partitioning by quarter makes perfect sense. If instead,
the archival is organized by year, then partitioning by year would be more relevant.

Partition Column Choice

The column chosen for the partition function should be the same column that the
columnstore index is ordered by. Partitioning and rowgroup elimination work best when
they operate on the same data sets and that can only be accomplished when they use
the same sort criteria. Therefore, if a columnstore index is ordered by a particular date
column, then partitioning should also be configured on that column. Partitioning by
another column may result in worse performance as SQL Server needs to scan more
rowgroups across more partitions to retrieve the data it needs.

Storage Choice
If data within a columnstore index is accessed differently, the data files used for the
partitioned table can mirror that usage. If older data is rarely accessed and can tolerate
more latency, then it can be moved onto slower/less expensive hardware. If newer data
needs to be highly available with minimal latency, then it can be placed on faster and
more expensive storage.
As a result, a large table can be split up to save money. Every terabyte of data that
lands on cheaper storage represents cost savings that can be easily quantified. While
partitioning’s primary purpose is not to save money, the ability to shift workloads via the
strategic placement of data files can achieve this without significant effort.
176
Chapter 11 Partitioning

When building a partitioned table, identify if the table has data that follows different
usage patterns and assign data files based on that usage to slower or faster storage,
when possible. If performance requirements change with time, data files can be moved
between storage locations to ensure that SLAs are still met, even if data that was once
rarely needed becomes critical to frequent analytic processes.

Additional Benefits
One of the greatest benefits of table partitioning is that it requires no application code
changes. All partitioning structures are internal to SQL Server and have no bearing
on an application beyond the performance experienced as the table is accessed. This
also means that partitioning can be tested out for analytic data and kept or rolled
back depending on the results of that testing. A “rollback” would consist of creating a
second copy of data and swapping it into the production location, but is a reasonable
process with analytic data where data load processes are well within the confines of data
architecture. This allows partitioning to be tested with minimal impact on the code that
developers write to consume this data.
Partitioning is an optional step when implementing a columnstore index, but can
improve maintenance, speed up analytic workloads, and potentially save money. This
feature should be targeted at larger tables with at least tens of millions of rows and that
are expected to grow rapidly over time.

177
CHAPTER 12

Nonclustered
Columnstore Indexes
on Rowstore Tables
Clustered columnstore indexes provide a primary storage mechanism for analytic data.
For tables that are intended for use as OLAP data sources, this is the optimal choice and
will provide a data structure that facilitates fast and efficient reads and writes against
analytic data.
For tables that are primarily transactional in nature, but also have analytic queries
executed against them, a clustered columnstore index is not appropriate. There are a
handful of options available to manage those additional analytic workloads, including

• Create covering rowstore indexes.

• Separate OLTP and OLAP.

• Create nonclustered columnstore indexes.

• Accept OLAP queries with no modifications to database schema.

Whereas a purely OLTP or OLAP workload can be managed with purely transactional
or analytic data structures, a mixed workload is more complex and requires a more
careful inspection of reads and writes to understand how to best manage it. This chapter
will discuss each alternative and when they are most appropriate, providing guidance on
how to choose the best storage methodology for a given workload.

179
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_12
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Use Rowstore Indexes

This is the typical startup default for most transactional applications. Data begins
small and grows over time. When row counts are low, running analytic queries against
transactional data is performant enough that no significant modifications are needed to
service them.
Relying on rowstore tables and single-column nonclustered rowstore indexes will
suffice until data size and contention become significant enough to result in
• Locking
• Blocking
• Excessive resource consumption
• Long-running queries
Once that threshold has been reached, another solution is required. Generally
speaking, running analytic processes against rowstore indexes will result in index scans,
heavy reads, and high memory usage, and those costs do not scale well as data grows
into millions or billions of rows. Figure 12-1 shows a simplification of the challenges of
running mixed workloads against a rowstore table.

Figure 12-1. The challenge of OLAP and OLTP queries against rowstore tables
While mixing analytic and transactional needs works for small, less busy tables,
challenges are guaranteed to arise when usage or size becomes significant. As a rule,
it is generally advisable to avoid relying on rowstore tables for analytics if the table is
expected to grow large.
Covering indexes allow for a rowstore index to completely cover an analytic query.
This can offer improvement from clustered index scans and key lookups and is a good
solution when analytics are limited in scope to a handful of well-understood queries.
For example, consider the query in Listing 12-1.
180
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Listing 12-1. Example Analytic Query Against a Transactional Table

SELECT
       COUNT(*) AS sale_count,
       COUNT(DISTINCT SalespersonPersonID) AS sales_people_count,
       SUM(CAST(IsUndersupplyBackordered AS INT)) AS
undersupply_backorder_count
FROM Sales.Orders
WHERE CustomerID = 90
AND OrderDate >= '1/1/2015'
AND OrderDate < '1/1/2016';

This query is a common example of analytics against transactional data. Figure 12-2
shows the execution plan for this query.

Figure 12-2. Execution plan for an analytic query against a transactional table

The execution plan shows an index seek that filters on the index for CustomerID. The
key lookup retrieves the other columns from the clustered index that are needed
to satisfy the query. The IO for the query as found in STATISTICS IO is shown in
Figure 12-3.

Figure 12-3. STATISTICS IO for an analytic query against a transactional table

181
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Note that the key lookup incurred quite a bit of logical IO as SQL Server had to
return to the clustered index to retrieve the OrderDate, SalespersonPersonID, and
IsUndersupplyBackordered columns. If this query is executed often and does not vary
significantly in form, a covering index can be an adequate way to manage it. Listing 12-2
creates an index that covers this sample query.

Listing 12-2. Nonclustered Rowstore Index That Fully Covers an Analytic Query

CREATE NONCLUSTERED INDEX NCI_Orders_covering

ON Sales.Orders (OrderDate, CustomerID)
INCLUDE (SalespersonPersonID, IsUndersupplyBackordered);

After creating this new covering index, the execution plan for the test query has been
simplified to what is seen in Figure 12-4.

Figure 12-4. Execution plan for an analytic query using a rowstore covering index

The execution plan shows an index seek, rather than a seek plus a key lookup, as was
the case before. The updated output of STATISTICS IO is shown in Figure 12-5.

Figure 12-5. STATISTICS IO for an analytic query using a rowstore covering index

The new STATISTICS IO output shows a seven times decrease in IO, from 470 reads
to 67. The covering index provides a significant performance boost to the analytic query,
and as an added bonus can reduce contention as the new index is a separate object that
can be read separately from the clustered index.
Covering indexes are great tools for analytic queries that are specific, consistent, and
few in number. There is a danger in creating excessive numbers of covering indexes to try

182
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

and keep up with new analytic queries. Over-indexing is a legitimate problem that can
hamper write performance, waste storage, and consume significant memory. Similarly,
covering indexes that contain too many columns are destined to waste space. For all
intents and purposes, a covering index with too many columns can quickly devolve into
a copy of the table, which is expensive to maintain.
The ideal scenarios where a covering index can be useful can be generalized as
follows:

• Very limited number of specific, targeted analytic queries.

• Analytic queries that do not change often, if ever.

• Analytic queries that occur often enough to warrant some indexing

action be taken.

• Column list is limited.

• Environment is conducive to removal of the index in the future, when

no longer needed.

Separate OLAP and OLTP Processes

One solution to the challenge of running analytics against transactional tables is to
separate OLAP data and processes from OLTP data and processes. There are many ways
to tackle this challenge, including

• AlwaysOn Availability Groups

• Replication

• Separate analytic tables, fed by ETL

• Third-party analytics tools

• Third-party hardware solutions

Figure 12-6 shows a simple representation of workloads against an OLAP data copy.

183
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Figure 12-6. OLAP workloads targeted at a data copy

While separating transactional and analytic data is by far the most effective way of
managing the challenge of mixed workloads, it is also the most disruptive.
AlwaysOn and Replication are mature tools that provide different ways to produce
secondary copies of data for use by analytic processes. They require new components
of SQL Server to be learned and implemented, as well as a commitment for additional
hardware to support a new target for copies of OLTP data. They can also provide high
availability, in the event that the primary transactional database fails.
If a database already uses AlwaysOn Availability Groups, then nonclustered
columnstore indexes can be leveraged to off-load analytic queries to a readable
secondary. Similarly, if Replication is used, the Replication target can be outfitted with a
clustered or nonclustered columnstore index, if its primary purpose is to service analytic
queries.
In addition to these built-in tools, an architect can create their own data copies and
manage them via SQL Server Integration Services or some other data movement process.
While manually building a data-copy process is a bit more labor intensive, it allows
for the use of existing tools and for any amount of customization. This is often how
organizations start their data warehousing environments.
Many third-party tools will perform similar tasks and allow for additional
processing/management of data. The upside to investing in a tool by an outside
organization is that the time, resources, and expertise needed to build and maintain it
fall squarely on their shoulders, freeing up resources for other projects. The downsides
are cost and investment in a new tool that may potentially be difficult to quit in the future
if needed.

184
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

While solutions such as these incur latency, the latency can be controlled via
configuration settings within each of them so that it remains within the bounds of
organizational needs.

Nonclustered Columnstore Indexes

Introducing other solutions to analytic queries against transactional data sets the stage
for how and when nonclustered columnstore indexes can effectively be used to manage
analytic queries.
Nonclustered columnstore indexes are architected similarly to clustered
columnstore indexes and exhibit similar behavior. Rowgroups, segments, and
compression may be analyzed in the same fashion as with clustered columnstore
indexes and best practices should be followed similarly. This means that they perform
exceptionally well for analytic-style queries that aggregate millions of rows in a single
operation. As with clustered columnstore indexes, though, DELETE and UPDATE
operations can incur fragmentation and waste space over time, and UPDATE operations
may perform poorly against large nonclustered columnstore indexes.
A huge difference between clustered and nonclustered columnstore indexes is that
the nonclustered version allows for a column list to be specified. While data order is
enforced by the rowstore clustered index, being able to pick and choose columns that are
included can isolate those needed for analytics as opposed to those used primarily for
transactional operations.
Note that the rowstore clustered index key columns are automatically included in
nonclustered columnstore indexes by default and cannot be removed. They are required
to support UPDATE and DELETE operations against existing rowgroups.
The primary benefits of nonclustered columnstore indexes are as follows:

• Allow real-time operational analytics.

• A column list can be specified.

• No app code changes are required.

• Columnstore index is stored on separate pages from rowstore

indexes, thereby reducing contention.

• Index may be filtered, assigned compression delay, or off-loaded to

an AlwaysOn replica.

185
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Like clustered columnstore indexes, nonclustered columnstore indexes require

some foresight to implement effectively. Because the clustered index is transactional in
nature and the table is likely the target of frequent small write operations, considerations
need to be made to the architecture of nonclustered columnstore indexes to ensure that
the impact of write operations is managed effectively.
Figure 12-7 illustrates a simplified illustration of workloads against a table with both
a clustered rowstore index and a nonclustered columnstore index.

Figure 12-7. Mixed workloads against rowstore and columnstore indexes

The takeaway from having both rowstore and columnstore indexes in a single table is
that both can be used simultaneously, allowing transactional and analytic operations to
execute side by side, rather than exclusively competing for resources.
To test the impact of nonclustered columnstore indexes on a transactional table,
consider the script in Listing 12-3.

Listing 12-3. Creation of a Nonclustered Columnstore Index on a Rowstore Table

CREATE NONCLUSTERED COLUMNSTORE INDEX NCCI_Orders ON Sales.Orders

(OrderDate, CustomerID, IsUndersupplyBackordered, SalespersonPersonID);

In the index creation statement, only columns involved in the analytic query were
included. More can be added if needed for other common queries, but the goal is to
ensure that analytics are used by the columnstore index without the need to refer to the
clustered rowstore index.
Once added, the test query in Listing 12-1 can be executed again. The execution plan
is shown in Figure 12-8.

186
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Figure 12-8. Execution plan for an analytic query using a nonclustered

columnstore index

The execution plan shows that the columnstore index was exclusively used to return
query results. Figure 12-9 shows the output of STATISTICS IO for the query.

Figure 12-9. STATISTICS IO for an analytic query using a nonclustered

columnstore index

Reads are about the same as with a covering index, with the added benefit that key
lookups to the clustered index are avoided, ensuring no contention with other queries
that are using it at the same time.

Managing Hot, Warm, and Cold Transactional Data

While nonclustered columnstore indexes can provide exceptional performance for
real-time operational analytics, they have one glaring weakness: write operations.
Transactional tables tend to incur small, incremental changes – inserts, updates, and
deletes that affect a few rows at a time and that occur frequently. Columnstore indexes
are designed for large bulk inserts, occasional deletes, and minimal updates.
Despite these immediate challenges, tools are available to manage these challenges
when implementing nonclustered columnstore indexes to help maintain optimal
performance.

187
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Compression Delay
Transactional data is often written and read frequently for a short span of time, after
which it is subject to less and less data access over time. The new data can be viewed as
hot data, whereas data that is somewhat old can be viewed as warm data, and older data
can be seen as cold data. While the table is transactional in nature, a vast majority of data
access will involve only the hot data.
For example, consider an online order tracking system. The following is a summary
of some of the steps involved in the order creation, processing, and completion, as well
as notes on data access:

1. Order is placed for some cool swag. Data is written to a database.

2. Order is processed. Data is updated in the database.

3. Item is retrieved and packaged. Data is updated in the database.

4. Item is shipped. Data is updated in the database.

5. Item is tracked until delivery. Data is updated in the database with

each shipping update.

6. Item is delivered. Data is updated in the database.

7. User leaves feedback on the order process and item.

8. User accesses this data read-only via the order history list in
the future.

9. Eventually, the user rarely/never accesses this data again.

This flow of data indicates a clear progression from hot data (#1–6) to warm data
(#7–8) to becoming cold data (#9). Analytic queries will most frequently address large
swaths of historical data, which is more apparent starting at step #6 on the list. Tracking
specific orders is undoubtedly a transactional need and applies most to steps #1–5 as
they correlate most with someone that wants to know where their order is and when it
will arrive.
A columnstore index will perform optimally the closer to the bottom of the list we
go, whereas the frequent updates at the top of the list will be burdensome to compressed
rowgroups. Each DELETE operation will flag a row as deleted in a rowgroup and waste
space. Each UPDATE operation will delete a row and also insert a new row into the
delta store.

188
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

To manage this challenge, compression delay can be used on a columnstore index.

This feature delays the movement of data from the delta store into compressed segments
by a preset amount of time. That timeframe is based on how long the data is expected
to still be hot. During that period of time where compression is delayed, there will be no
need to worry about the impact of DELETE or UPDATE operations against compressed
rowgroups. Conversely, reads against the uncompressed delta store data will be slower
than when the data ultimately is compressed. For a large table, though, the percentage of
data that is hot will be small, whereas the amount of data that is warm or cold will comprise
most of the table. Equally important is that the clustered rowstore index will service
transactional queries, allowing data in the delta store to be avoided by those processes.
Consider the example of a columnstore index on Sales.Orders from Listing 12-3. If
there typically was a period of 10 minutes in which the data was frequently updated,
after which the data began to cool off and see less modifications, compression delay
could be added to the index creation statement, as seen in Listing 12-4.

Listing 12-4. Nonclustered Columnstore Index Creation with

Compression Delay
CREATE NONCLUSTERED COLUMNSTORE INDEX NCCI_Orders
ON Sales.Orders (OrderDate, CustomerID, IsUndersupplyBackordered,
SalespersonPersonID)
WITH (COMPRESSION_DELAY = 10 MINUTES);

The index will initially perform exactly as it did before. The difference will be when
data is written to the delta store. Any rows inserted into the delta store will be retained
for at least 10 minutes prior to writing to compressed rowgroups. Without compression
delay, the tuple mover will periodically move rows from the delta store to compressed
rowgroups With this feature, it will now wait at least 10 minutes before doing so.
The compression delay may be adjusted on an existing index, if needed, without the
need for a rebuild. For example, if it is determined that data is hot for 30 minutes, rather
than 10, an ALTER INDEX statement can adjust the compression delay accordingly, as
seen in Listing 12-5.

Listing 12-5. Using an ALTER INDEX Statement to Adjust Compression Delay

ALTER INDEX NCCI_Orders ON Sales.Orders

SET (COMPRESSION_DELAY = 30 MINUTES);

189
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Configuring compression delay is a balance between avoiding excessive writes

against compressed rowgroups while ensuring that the delta store does not become
so large that it adversely affects analytics. Ideally, rowstore indexes will service the
transactional queries that operate on hot data, whereas the columnstore index handles
analytic queries against larger volumes of data (that is mostly warm or cold).
Testing is necessary when determining the ideal value for compression delay, but the
following is a set of guidelines to assist in finding a good starting point:

1. Measure new rows inserted over time. Ideally, the delta store
would contain no more than a few rowgroups worth of data (a few
million or less) that are not used heavily for analytics.

2. Measure update and delete operations and the period of time

in which they most frequently occur. Try to set the compression
delay to encompass that period of time.

3. If analytics do not target new data, then a higher compression

delay is acceptable. Otherwise, it may need to be lowered to avoid
frequent scans of the delta store.

4. If data load processes modify data heavily as part of insert

operations, ensure that the compression delay period is long
enough to encompass a typical data load. Alternatively, modify
the data load process to use staging/temporary tables to modify
data prior to its final insertion.

Compression delay can be a large or small number. If a transactional table has 1,000
new rows inserted per hour and those rows are heavily modified for 6 hours, then a
compression delay of 360 minutes would be perfectly acceptable. In that scenario, the
delta store would contain an average of 6,000 rows, which is small enough that scanning
it would not be painful.
Alternatively, if a table has 2,000,000 rows inserted per hour, then compression delay
would need to be more restrictive. If data in that table is hot for 2 hours, then an ideal
compression delay value would be somewhere in the range of 30–120 minutes. Testing
would be required to balance the needs of write processes vs. those of analytic read
processes. The diagram in Figure 12-10 shows how compression delay impacts the flow
of data for transactional and analytic queries with a nonclustered columnstore index.

190
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Figure 12-10. Data flow for OLAP and OLTP queries against a nonclustered
columnstore index with compression delay

The goal represented in this data flow is for transactional queries to target the
rowstore indexes and the delta store, whereas analytic queries target the columnstore
index. Compression delay helps to ensure that OLTP writes are not impacting the
performance of the compressed rowgroups within the nonclustered columnstore index.
OLTP reads would be expected to target the rowstore indexes and would rarely make use
of the columnstore index.
Compression delay can be applied to clustered columnstore indexes as well using
the same syntax. This can be an excellent solution when data load processes fulfill one of
these criteria:

• They have an overall long duration.

• Data is trickle loaded, rather than bulk loaded.

• Updates target data after it is initially inserted.

• A primarily analytic table is subject to OLTP queries as part of the

data load process.

The query in Listing 12-6 alters the clustered columnstore index in

WideWorldImportersDW to have a 60 minute compression delay.

191
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Listing 12-6. Using an ALTER INDEX Statement to Adjust Compression Delay on

a Clustered Columnstore Index
ALTER INDEX CCI_fact_sale_CCI ON Fact.Sale_CCI
SET (COMPRESSION_DELAY = 60 MINUTES);

A 60 minute compression delay on a clustered columnstore index would provide a

60 minute window in which a data load process could freely insert data in any fashion,
and then update or delete from it if needed. Since the data resides in the delta store,
the additional updates and deletes would not impact compressed rowgroups, therefore
removing the introduction of fragmentation by the data load process. In addition, poor
performance stemming from UPDATE queries would be avoided. After the 60 minute
period has passed, the data would begin to be inserted into the columnstore index.
Assuming the data load process has completed its data modifications against that data,
then it would be inserted once and then left alone in its pristine compressed state.
The compression delay settings for columnstore indexes in a database can be
retrieved using a query against sys.indexes, as shown in Listing 12-7.

Listing 12-7. Query to Return Index Compression Delay Information

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       indexes.type_desc AS index_type,
       indexes.compression_delay
FROM sys.indexes
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
WHERE indexes.type_desc IN ('NONCLUSTERED COLUMNSTORE', 'CLUSTERED
COLUMNSTORE');

The column compression_delay provides a number (in minutes) indicating the

current compression delay setting for that index. Zero indicates that no compression
delay is used for an index. Note that the query filters exclusively on columnstore indexes.
Other index types will have NULL entered for compression delay as it is not an option for
any other index types aside from columnstore indexes.
Figure 12-11 shows the results of this query and the compression delay for each
columnstore index in the database.
192
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Figure 12-11. Compression delay for each columnstore index in a database

Row 4 (circled) shows the index that was altered by the query in Listing 12-6 and
confirms the compression delay setting of 60 minutes.
Fragmentation due to rows being deleted can be quantified using the percentage of
rows being deleted, using the query shown in Listing 12-8.

Listing 12-8. Query to Return Index Fragmentation Due to Delete Operations

SELECT
       objects.name,
       partitions.partition_number,
       dm_db_column_store_row_group_physical_stats.row_group_id,
       dm_db_column_store_row_group_physical_stats.total_rows,
       dm_db_column_store_row_group_physical_stats.deleted_rows,
       CAST(100 * CAST(deleted_rows AS DECIMAL(18,2)) / CAST(total_rows AS
DECIMAL(18,2)) AS DECIMAL(18,2)) AS percent_deleted
FROM sys.dm_db_column_store_row_group_physical_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_physical_stats
.object_id
INNER JOIN sys.partitions
ON partitions.object_id = objects.object_id

193
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

AND partitions.partition_number = dm_db_column_store_row_group_physical_

stats.partition_number
AND partitions.index_id = dm_db_column_store_row_group_physical_stats
.index_id
WHERE objects.name = 'Sale_CCI'
ORDER BY dm_db_column_store_row_group_physical_stats.row_group_id;

This query specifically targets a single table (Sale_CCI) and returns the total rows,
deleted rows, and percent deleted per rowgroup, as seen in Figure 12-12.

Figure 12-12. Fragmentation due to deleted rows in a columnstore index

In this example, fragmentation is approximately 5% across the index. Generally

speaking, when deleted rows comprise more than 10% of a columnstore index, it may be
considered fragmented enough to warrant further actions to be taken in order to remove
and/or prevent further fragmentation.
Compression delay is an exceptionally useful setting when creating nonclustered
columnstore indexes that can offset heavy transactional write operations against a
table during the period of time when data is still hot. In addition to improving write
performance, compression delay will reduce fragmentation, improve compression
efficiency, and decrease memory consumption for the index.

194
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Filtered Nonclustered Columnstore Indexes

Like nonclustered rowstore indexes, nonclustered columnstore indexes can have filters
applied to them when created. This allows the columnstore index to be targeted, focused
on cold and warm data while avoiding indexing hot data. These filters will often target
columns that enumerate statuses, workflows, and location within the data life cycle.
Consider the example of an order processing system. In this system, orders are
frequently manipulated up until a certain point in the process. Maybe it is when the
order is picked, shipped, or received, but at some point in its life cycle, the data shifts
from hot data to warm data and then later to cold data.
A filtered index can be used to target a nonclustered columnstore index to include
only data that is no longer hot. For example, consider the index creation statement in
Listing 12-9.

Listing 12-9. Creation of a Filtered Nonclustered Columnstore Index

CREATE NONCLUSTERED COLUMNSTORE INDEX NCCI_Orders

ON Sales.Orders (OrderDate, CustomerID, IsUndersupplyBackordered,
SalespersonPersonID)
WHERE PickedByPersonID IS NOT NULL;

This index will only include rows that happen to have a value for the
PickedByPersonID column. As a result, rows will not be inserted into the columnstore
index until they have been boxed and are ready to ship. Because of this, any data
manipulation that occurs prior to this point in time will not impact the performance of
the columnstore index.
Listing 12-10 calculates the number of rows that meet the filter criteria vs. the
number that do not.

Listing 12-10. Calculating the Number of Rows That Meet the Filtered Index
Criteria
SELECT
SUM(CASE WHEN PickedByPersonID IS NULL THEN 1 ELSE 0 END) AS
orders_not_picked,
SUM(CASE WHEN PickedByPersonID IS NOT NULL THEN 1 ELSE 0 END) AS
orders_picked
FROM Sales.Orders
195
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

The results in Figure 12-13 show that about 14% of the rows in the table are not
picked and therefore are still classified as hot data.

Figure 12-13. Count of rows that meet the filtered index criteria

Typically, hot data will constitute a small fraction of data in the table. If a simple
filter can be used to determine whether data is hot or not, then it can be applied to
the nonclustered columnstore index to ensure that its performance is not negatively
impacted by it.
Consider the analytic query presented in Listing 12-1. If this query was intended
to only target data that is no longer hot, then adding the filter criteria used on the
nonclustered columnstore index to this query would allow it to use the filtered version of
the index, as seen in Listing 12-11.

Listing 12-11. Query That Uses a Filtered Nonclustered Columnstore Index

196
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

The execution plan for this query is shown in Figure 12-14.

Figure 12-14. Execution plan using a filtered nonclustered columnstore index

As expected, the filtered nonclustered columnstore index was used to execute

the query.
If needed, the query in Listing 12-12 shows all indexes in a database that have a filter
defined on them, which can be used to review columnstore indexes and their details.

Listing 12-12. Query That Lists Indexes with Filters in a Given Database

SELECT
       indexes.name,
       indexes.type_desc,
       indexes.filter_definition
FROM sys.indexes
WHERE indexes.has_filter = 1;

The results of the query are useful in understanding what filters exist and their
definitions, as shown in Figure 12-15.

Figure 12-15. List of indexes with filters defined on them

If the database contains many filtered indexes, the query could be further refined to
omit rowstore indexes or other noise.

197
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

There is a single pitfall in filtered indexes that needs to be reviewed prior to using
them, and that is filter column sensitivity. Ideally, as data moves from hot to warm
to cold, it is added to the filtered nonclustered columnstore index and rarely (if ever)
removed. If rows are capable of alternating between satisfying the filter clause and then
no longer meeting the criteria, there is a danger that a single row in a table could be
added and removed from the columnstore index many times. If this is common, then the
filtered index may end up being as fragmented as a nonfiltered index would be.
Consider a scenario using the filtered nonclustered columnstore index created in
Listing 12-9. If a row were assigned a value for PickedByPersonID, it would immediately
be inserted into the delta store for the columnstore index where it would await
compression via the tuple mover. If PickedByPersonID is then reset to NULL, it would
be deleted from the columnstore index as it no longer meets the filter criteria. When
designing any filtered index, there is value in ensuring that rows will not move in and out
of the index often and that data follows a one-way journey from hot to warm to cold.
Compression delay can be combined with a filtered nonclustered columnstore index
as a way to buffer out-of-band changes to data that cause it to move between hot and
warm via routine processes. Listing 12-13 shows a filtered nonclustered columnstore
index that includes compression delay.

Listing 12-13. Creation of a Filtered Nonclustered Columnstore Index with

Compression Delay
CREATE NONCLUSTERED COLUMNSTORE INDEX NCCI_Orders
ON Sales.Orders (OrderDate, CustomerID, IsUndersupplyBackordered,
SalespersonPersonID)
WHERE PickedByPersonID IS NOT NULL
WITH (COMPRESSION_DELAY = 30 MINUTES);

This index provides additional buffering against the possibility that rows that have
been assigned a PickedByPersonID might be
• Deleted
• Updated
• Have PickedByPersonID set back to NULL
A transactional table may make use of filtered nonclustered columnstore indexes,
compression delay, or both and do so with success. The details as to which features to
use and how to configure them should be driven by organizational need and its data life
198
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

cycle. There is no one-size-fits-all solution in making these decisions. One table in one
database for organization may benefit from a compression delay of 10 minutes, whereas
1,440 minutes might be optimal for another table. Filters might fully encapsulate a key
use case in one workflow when others may be unable to easily be filtered for in an index.

C
ode Changes
If compression delay and filtered nonclustered columnstore indexes are unable
to effectively manage analytic workloads against an OLTP table, it is possible that
organizational logic, code, or both require review.
If application or database code can be improved in a way that supports both analytic
and transactional workloads against the same table, then making those changes can
allow an organization to prevent having to architect a more complex and expensive
OLAP solution.
Ultimately, a solution exists for any analytic data challenge, but the ideal solution
will be the one that costs the fewest resources while being capable of scaling most easily
into the future. Real-time operational analytics are not appropriate for all workloads, and
there will often be scenarios where a more expensive and resource-intensive solution
(such as AlwaysOn or Replication) is the correct solution.

ertipaq Optimization for Nonclustered

V
Columnstore Indexes
In Chapter 5, columnstore compression was discussed in detail. An important issue
that arose was that when a nonclustered rowstore index was added to a clustered
columnstore index, rows that were inserted in the future could not benefit from Vertipaq
optimization. Vertipaq optimization reorders rows, while SQL Server compresses a
rowgroup, allowing like values to be grouped together and therefore compress more
effectively. Without this feature available, newly inserted data would not compress as
efficiently and both storage space and memory would be wasted over time.
A logical question to ask is, “Do nonclustered columnstore indexes benefit from
Vertipaq optimization?” To test this and determine if Vertipaq optimization will be
used or not, new rows will be inserted into Sales.Orders, as provided by the query in
Listing 12-14.

199
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Listing 12-14. Inserting Rows into Sales.Orders to Test Vertipaq Optimization

INSERT INTO Sales.Orders

(   OrderID, CustomerID, SalespersonPersonID, PickedByPersonID,
ContactPersonID, BackorderOrderID, OrderDate, ExpectedDeliveryDate,
CustomerPurchaseOrderNumber, IsUndersupplyBackordered, Comments,
DeliveryInstructions, InternalComments, PickingCompletedWhen,
LastEditedBy, LastEditedWhen)
SELECT
    73595 + ROW_NUMBER() OVER (ORDER BY OrderID) AS OrderID,
    CustomerID,
    SalespersonPersonID,
    PickedByPersonID,
    ContactPersonID,
    BackorderOrderID,
    OrderDate,
    ExpectedDeliveryDate,
    CustomerPurchaseOrderNumber,
    IsUndersupplyBackordered,
    Comments,
    DeliveryInstructions,
    InternalComments,
    PickingCompletedWhen,
    LastEditedBy,
    LastEditedWhen
FROM Sales.Orders;

This insert operation doubles the size of the table. Once complete, the query in
Listing 12-15 can be executed to confirm the current state of rowgroups within the
nonclustered columnstore index.

Listing 12-15. Query to Display Columnstore Rowgroup Metadata After a Large

INSERT Operation
SELECT
objects.name AS table_name,
partitions.partition_number,

200
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

       dm_db_column_store_row_group_physical_stats.row_group_id,

dm_db_column_store_row_group_physical_stats.has_vertipaq_
optimization,
       dm_db_column_store_row_group_physical_stats.state_desc
FROM sys.dm_db_column_store_row_group_physical_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_physical_stats.
object_id
INNER JOIN sys.partitions
ON partitions.object_id = objects.object_id
AND partitions.partition_number = dm_db_column_store_row_group_physical_
stats.partition_number
AND partitions.index_id = dm_db_column_store_row_group_physical_stats.
index_id
WHERE objects.name = 'Orders'
ORDER BY dm_db_column_store_row_group_physical_stats.row_group_id;

The results in Figure 12-16 show both the existing compressed rowgroup and the
newly created open rowgroup.

Figure 12-16. Columnstore metadata after a large INSERT operation

Data in the new rowgroup still resides in the delta store as a compression delay of
30 minutes was specified when the columnstore index was created. To avoid waiting 30
minutes for the rowgroup to be processed by the tuple mover, the T-SQL in Listing 12-16
will be used to enable that process immediately.

Listing 12-16. T-SQL to Enable the Tuple Mover to Process Open Delta
Rowgroups
ALTER INDEX NCCI_Orders ON sales.Orders REORGANIZE WITH
(COMPRESS_ALL_ROW_GROUPS = ON);

201
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

The results in Figure 12-17 show the state of the columnstore index immediately
after this script completed.

Figure 12-17. Columnstore metadata after the open rowgroup is processed

Note that the newly created rowgroup (row_group_id = 2) is compressed and is using
Vertipaq optimization. If the metadata script in Listing 12-15 is executed again after a few
minutes pass, then the rowgroup in the TOMBSTONE state will be removed, as shown in
Figure 12-18.

Figure 12-18. Columnstore metadata after the tombstone rowgroup is

automatically cleaned up

The result of the INSERT operation is 2 rowgroups, one created alongside the initial
creation of the columnstore index and the other created when a new set of rows were
inserted and compressed. As shown in this example, Vertipaq optimization will be used
for nonclustered columnstore indexes. This is a boon to columnstore compression and
ensures that each segment is compressed as efficiently as possible, despite representing
transactional data that may be subject to change more frequently than typical
analytic tables.
It is the inclusion of the clustered rowstore index key columns in the nonclustered
columnstore index that enables Vertipaq optimization to be used. Without the key
columns included, linking rows within the columnstore index back to the rowstore
index would be impossible and the columnstore index would need to be rebuilt
whenever data is modified. Such a cost would be prohibitively high. Having clustered

202
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

key columns available ensures that rows within the nonclustered columnstore index
can be freely reordered without any danger that they could not easily be tied back to the
clustered index.

Testing Nonclustered Columnstore Indexes

Because mixing analytic and transactional workloads incurs risk, testing is a key
component of implementing nonclustered columnstore indexes. The following is a short
list of guidelines to assist in testing nonclustered columnstore indexes.

• Test trickle inserts. Does inserting lots of single rows or small

batches of rows perform similarly before and after the nonclustered
columnstore index was added?

• Test common updates. Are UPDATE operations used against hot

data? If so, test their performance before and after the columnstore
index addition.

• Test compression delay. For OLTP rowstore tables, compression

delay is likely a beneficial option to implement. Test different values
of compression delay to find a sweet spot where write operations are
as efficient as possible while generating little fragmentation in the
columnstore index.

• Test filters. If a common delineation of hot data is defined by a

column or set of columns, then adding a filter to the index can help
isolate hot data from the data typically used for analytics. Test these
carefully to ensure that analytic queries can use the index and that
the filter columns do not get updated often.

• Perform a before and after load test. When in doubt, a general

load test of common operations that compares performance
before and after the nonclustered index is added can be helpful in
understanding the impact of the new index. Ideally, a nonclustered
columnstore index should greatly improve analytic performance
while not negatively impacting write operations.

203
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

Nonclustered columnstore indexes provide an excellent tool for mixed workloads

where transactional and analytic queries target a single table or set of tables. If the
expense of separating OLAP from OLTP is too high, then this can be a cost-effective way
to manage common analytic queries without the need to make significant architectural
changes or change application code.

Don’t Forget to Drop Unneeded Indexes!

Nonclustered columnstore indexes will often be implemented to replace one or many
nonclustered rowstore indexes that previously serviced analytic queries. When this is the
case, be certain to test and drop the old covering indexes when they are determined to
no longer be needed.
If a nonclustered columnstore index can replace many covering rowstore indexes, then
its addition may very well save computing resources after all of the old, unneeded indexes
are dropped. The query in Listing 12-17 returns index usage statistics for a given table.

Listing 12-17. Script That Returns Index Usage Data for One Table

SELECT
       tables.name AS TableName,
       indexes.name AS IndexName,
       dm_db_index_usage_stats.user_seeks,
       dm_db_index_usage_stats.user_scans,
       dm_db_index_usage_stats.user_lookups,
       dm_db_index_usage_stats.user_updates,
       dm_db_index_usage_stats.last_user_seek,
       dm_db_index_usage_stats.last_user_scan,
       dm_db_index_usage_stats.last_user_lookup,
       dm_db_index_usage_stats.last_user_update
FROM sys.dm_db_index_usage_stats
INNER JOIN sys.tables
ON tables.object_id = dm_db_index_usage_stats.object_id
INNER JOIN sys.indexes
ON indexes.object_id = dm_db_index_usage_stats.object_id
AND indexes.index_id = dm_db_index_usage_stats.index_id
WHERE tables.name = 'Orders'

204
Chapter 12 Nonclustered Columnstore Indexes on Rowstore Tables

The results are provided in Figure 12-19.

Figure 12-19. Index usage statistics output

The data returned by sys.dm_db_index_usage_stats is reset when SQL Server restarts,

so it is important to either persist it somewhere permanent or take regular snapshots
of it. Using the information returned, it can be quickly determined how often an index
is used for reads (seeks, scans, lookups) vs. writes (updates). An index with zero or few
reads over a long period of time can be safely dropped. Indexes with more frequent reads
would need additional review to determine the impact of dropping them on the queries
that regularly use them.
To summarize, the goal of secondary indexes on a clustered rowstore table is to
produce the simplest, most compact and efficient solution possible. For analytic queries
with a predictable and often repeated structure, nonclustered covering rowstore indexes
an be sufficient to provide good performance with minimal change.
If analytic needs against rowstore tables are more diverse in structure, then a single
nonclustered columnstore index can be a far more efficient way to cover them than
creating many nonclustered rowstore indexes.
If filters can be applied to any secondary or covering indexes (rowstore or
columnstore), then doing so can reduce their size and improve performance further by
avoiding hot data that is written to often.

205
CHAPTER 13

Nonclustered Rowstore
Indexes on Columnstore
Tables
Clustered columnstore indexes provide effective enough compression and data access
speeds that most typical analytic workloads will not require any other indexes to provide
adequate performance.
The key to this performance lies in the underlying data order for a given analytic
table. As long as queries using that data are able to filter based on the dimension that
the data is ordered by (typically time), they can benefit from rowgroup elimination. For
unusual queries that slice the data by other dimensions, the result will often be a full
scan of the columnstore index. For large tables, this is prohibitively expensive and will
necessitate another solution to help in optimizing these workloads.

Using Nonclustered Rowstore Indexes

Chapter 10 reviewed in detail the importance of data order in allowing for rowgroup
elimination. When data is ordered by a dimension and analytic queries filter on that
dimension, SQL Server can use columnstore metadata to automatically filter out
unneeded rowgroups.
Nonclustered rowstore indexes can be created on clustered columnstore tables
to provide additional support for common queries that do not follow along with the
ordering of the data. Care must be taken when adding supporting indexes to a clustered
columnstore index. The remainder of this chapter will discuss how nonclustered
rowstore indexes can be best used when needed, as well as best practices for
maintaining them.
207
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_13
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Consider a clustered columnstore index on the table Fact.Sale_CCI_Ordered that

is ordered by the column Invoice Date Key. This table is architected for filtering and
aggregation across the time dimension provided by the column Invoice Date Key. The
query in Listing 13-1 provides an analytic query that takes advantage of this natural
data order.

Listing 13-1. Query That Filters and Aggregates Across the Time Dimension of
an Ordered Clustered Columnstore Index
SELECT
COUNT(*),
SUM(Quantity) AS total_quantity
FROM Fact.Sale_CCI_ORDERED
WHERE [Invoice Date Key] >= '11/1/2015'
AND [Invoice Date Key] < '1/1/2016';

When executed, the result is returned quite quickly. Figure 13-1 shows the output of
STATISTICS IO.

Figure 13-1. STATISTICS IO output for a query against an ordered clustered

columnstore index

Note that the segment report shows 4 segments read and 21 skipped. In addition, the
logical reads are quite low given that this table contains 25 million rows. Both of these
are indicators that rowgroup elimination was used effectively and helped ensure that
only a small slice of the table needed to be read in order to process the query and return
results.
So long as the analytic queries line up filters, ordering, and aggregation with the
column(s) that the table is ordered by, then performance like this is to be expected.
Consider the very different query provided in Listing 13-2.

208
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Listing 13-2. Analytic Query That Does Not Use a Table’s Natural Ordering

SELECT
COUNT(*),
SUM(Quantity) AS total_quantity
FROM Fact.Sale_CCI_ORDERED
WHERE [Stock Item Key] = 186;

This query calculates quantity using Stock Item Key as the filter. Since the table
is ordered by Invoice Date Key, SQL Server has no natural way to achieve rowgroup
elimination. Figure 13-2 shows the STATISTICS IO output for this query.

Figure 13-2. STATISTICS IO output for a query against an unordered column in

an ordered clustered columnstore index

This time, all rowgroups were read in the columnstore index and none were skipped.
This performance can be attributed to there being a total of 227 values for Stock Item Key
in the table, but those values being scattered across all rowgroups without any particular
order. Listing 13-3 provides the query to return metadata about the Stock Item Key
column contents in the columnstore index.

Listing 13-3. Query to Return Metadata About the Stock Item Key Column

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       columns.name AS column_name,
       partitions.partition_number,
       column_store_segments.segment_id,
       column_store_segments.min_data_id,
       column_store_segments.max_data_id,
       column_store_segments.row_count

209
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id
INNER JOIN sys.indexes
ON indexes.index_id = partitions.index_id
AND indexes.object_id = partitions.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.columns
ON tables.object_id = columns.object_id
AND column_store_segments.column_id = columns.column_id
WHERE tables.name = 'Sale_CCI_ORDERED'
AND columns.name = 'Stock Item Key'
ORDER BY tables.name, columns.name, column_store_segments.segment_id;

Figure 13-3 provides the output of this query, showing the minimum and maximum
values for Stock Item Key for each rowgroup.

Figure 13-3. Columnstore segment metadata for the Stock Item Key columns

It is readily apparent when viewing this segment metadata that each rowgroup has
similarly wide ranges of values for Stock Item Key, and therefore SQL Server is unable to
effectively use metadata to perform much filtering using this column. It is worth noting
that a query for a value greater than 219 would allow for some rowgroup elimination as
the max_data_id for some rowgroups is 219, rather than 227 or 225. Ignoring that small
detail, it is safe to say that most queries that filter on Stock Item Key will be forced to scan
all rowgroups in the columnstore index in order to return results.
If queries that filter on Stock Item Key are frequent and their current performance is
unacceptably slow, then one solution to this challenge is to implement a nonclustered
rowstore covering index on top of the columnstore index. To cover the query presented
210
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

here, it is necessary to order on Stock Item Key and include Quantity. The query in
Listing 13-4 creates this index.

Listing 13-4. Query to Create a Covering Rowstore Index

CREATE NONCLUSTERED INDEX IX_Sale_CCI_ORDERED

ON Fact.Sale_CCI_ORDERED ([Stock Item Key]) INCLUDE (Quantity);

The index creation query takes about 45 seconds to complete. Once complete,
executing the sample query in Listing 13-2 produces a new execution plan and new IO,
as seen in Figure 13-4.

Figure 13-4. Query execution plan and IO when using a covering index

SQL Server exclusively uses the covering index to execute the query, resulting
in 394 reads and an index seek against the new index. This IO is slightly higher than
the previous example, but that should not be the sole decision-making metric when
determining indexing strategy. The query executes exceptionally fast as well. There are a
few key benefits to using a nonclustered rowstore index:

• Fast, predictable execution times, regardless of query parameters.

• Query reads against the covering index do not create contention with
queries against the clustered columnstore index.

• Predictable execution plans and IO.

211
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

The benefits of covering nonclustered rowstore indexes are found exclusively on

the reads. There are downsides to creating this index that must be understood before
deciding to create a covering index like this:

• Write times are slowed down as the nonclustered index needs to be

populated and mapped to the clustered columnstore index.

• Vertipaq optimization is not used when any nonclustered rowstore

indexes are present. The mapping process to the clustered
columnstore index makes reordering rows too inefficient of a process
to execute on the fly.

• Nonclustered rowstore index sizes will be inflated. Data from the

clustered index is duplicated within the nonclustered indexes. In
addition, the key that links the rowstore index back to the clustered
columnstore index is an 8-byte unique value that combines the
4-byte rowgroup ID and a 4-byte offset. If the nonclustered index
is not unique, additional (variable) storage will be used to enforce
uniqueness internally at the leaf level of the index.

Adding a nonclustered rowstore index to a clustered columnstore index is by no

means an inexpensive decision. Its price is storage space, memory usage, and write
speeds. Because of this, careful consideration must be taken prior to adding supporting
indexes to a clustered columnstore index.

E nforcing Constraints
One common use of nonclustered rowstore indexes on clustered columnstore indexes
is not performance related. Nonclustered rowstore indexes may be used to enforce
uniqueness on a table, either via a primary key or by using a unique nonclustered index
definition.
Creating unique constraints is no different than on a clustered rowstore table.
Listing 13-5 shows the create statement for an existing nonclustered primary key over a
clustered columnstore index.

212
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Listing 13-5. Creating a Unique Nonclustered Rowstore Index on Top of a

Clustered Columnstore Table
ALTER TABLE Fact.Sale ADD CONSTRAINT PK_Fact_Sale PRIMARY KEY NONCLUSTERED
( [Sale Key] ASC,
[Invoice Date Key] ASC);

This new index serves as a primary key for the underlying table, both enforcing
uniqueness and allowing for foreign keys to be created on other tables that reference
this table. For a scenario where either of these are important architectural requirements,
then creating a primary key on the clustered columnstore index is a good solution to this
problem.
The table Fact.Sale contains many nonclustered rowstore indexes defined on it. The
query in Listing 13-6 returns the index size for each index on the table.

Listing 13-6. Query to Return Index Size for Each Index on Fact.Sale

SELECT
    indexes.name AS Index_Name,
    SUM(dm_db_partition_stats.used_page_count) * 8 Index_Size_KB,
    SUM(dm_db_partition_stats.row_count) AS Row_Count
FROM sys.dm_db_partition_stats
INNER JOIN sys.indexes
ON dm_db_partition_stats.object_id = indexes.object_id
AND dm_db_partition_stats.index_id = indexes.index_id
INNER JOIN sys.tables
ON tables.object_id = dm_db_partition_stats.object_id
INNER JOIN sys.schemas
ON schemas.schema_id = tables.schema_id
WHERE schemas.name = 'Fact'
AND tables.name = 'Sale'
GROUP BY indexes.name
ORDER BY indexes.name

213
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

The results of the index size query are shown in Figure 13-5.

Figure 13-5. Index size details for one table

In addition to the clustered columnstore index and the nonclustered rowstore

primary key, there are another seven nonclustered rowstore indexes that each index a
foreign key column. Note that the index size for each supporting index is close to the
total size of the columnstore index itself. As a result, the table itself contains about 5.7MB
of compressed columnstore data, but the total space used for data and index space is
about 45MB.
This is an excellent example of an over-indexed table. If users truly search by that
many dimensions without regard for a single data order, then a columnstore index might
not be the ideal solution for this table. Alternatively, removing most of the extra indexes
may still allow for performance that is fast enough, even in their absence.
Using a nonclustered rowstore index to enforce uniqueness on a clustered columnstore
index can be a valuable tool when looking to enforce data integrity on OLAP data. It is
generally inadvisable to implement multiple rowstore indexes on a clustered columnstore
table unless the organizational need for those indexes greatly exceeds the cost.

Filtered Nonclustered Rowstore Indexes

One way to mitigate the cost of a nonclustered rowstore index on a columnstore index is
to add a filter. Filtered indexes target a smaller portion of the table, consume less storage
and memory, and can have a reduced impact on the compression and performance of
the columnstore index.
214
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Oftentimes, OLTP queries against an analytic table may target only a portion of the
data – maybe newer, older, or based on some other useful filter. Consider the T-SQL that
was tested in Listing 13-2. This query filters on Stock Item Key, which is not the column
that then the columnstore index was ordered on. To mitigate the cost of this query, a
nonclustered rowstore index was added, as was shown in Listing 13-4. What if this query
never targeted new data, but exclusively was used to analyze older sales? Consider the
updated index in Listing 13-7.

Listing 13-7. Example of a Filtered Nonclustered Rowstore Index on a Clustered

Columnstore Index
CREATE NONCLUSTERED INDEX IX_Sale_CCI_ORDERED
ON Fact.Sale_CCI_ORDERED ([Stock Item Key], [Invoice Date Key]) INCLUDE
(Quantity)
WHERE [Invoice Date Key] <= '1/1/2016';

When a query is executed that uses this index and filters for data prior to 1/1/2016,
it will use the filtered index rather than the columnstore index. Because the index is
smaller, the number of rows that are read and the IO will be less than for the unfiltered
version of this index.
One significant bonus of filtering a nonclustered rowstore index is that it allows
the index to target more specifically hot vs. warm vs. cold data. This provides multiple
benefits:

• Allows an index to target a specific use case, rather than all data

• Allows for better management of Vertipaq optimization

• Can facilitate the addition of multiple nonclustered indexes with less

impact on write performance or resource consumption

If there is a need for secondary indexing on a clustered columnstore index, consider

if filters can be added to reduce the size and impact of the nonclustered indexes. It’s an
easy way to improve performance and reduce storage and memory footprint, assuming
queries that use the index can use the filters consistently.

215
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Enabling Vertipaq Optimization

A downside of adding nonclustered rowstore indexes to a clustered columnstore index
is that SQL Server will automatically forgo Vertipaq optimization. This will result in
columnstore compression that is less effective than it could be if the feature were
used. Note that Vertipaq optimization is only skipped for new data that is inserted into
the table. Any data that existed in the columnstore index prior to the creation of the
nonclustered indexes will benefit from the improved performance granted by it.
If a nonclustered rowstore index is needed on a clustered columnstore index,
consider one of the following solutions that will allow for Vertipaq optimization to be
maintained on compressed columnstore segments.

Add Filters to Nonclustered Rowstore Indexes

In addition to targeting relevant data when needed, filters can allow for hot data to be
skipped when creating nonclustered indexes. This allows new data to be inserted into
the clustered columnstore index using Vertipaq optimization, whereas older data can
still be read using the nonclustered rowstore index.

Perform Periodic Index Maintenance

If nonclustered indexes are needed and filters cannot be applied, consider periodic
maintenance in which the following steps are taken:

1. Drop nonclustered rowstore indexes.

2. Rebuild the clustered columnstore index.

3. Re-create the nonclustered rowstore indexes.

This can be time-consuming, especially for a larger table, but can allow columnstore
indexes to be periodically rebuilt to be smaller and more efficient than before.
Note that partitioning can be a valuable tool here as index maintenance can target
specific partitions. The query in Listing 13-8 shows how partitions can be filtered to show
only those that contain rowgroups lacking Vertipaq optimization.

216
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Listing 13-8. Query to Return Rowgroups That Are Not Benefitting from Vertipaq
Optimization
SELECT DISTINCT
       objects.name,
       partitions.partition_number,
       dm_db_column_store_row_group_physical_stats.row_group_id,
       dm_db_column_store_row_group_physical_stats.has_vertipaq_
optimization
FROM sys.dm_db_column_store_row_group_physical_stats
INNER JOIN sys.objects
ON objects.object_id = dm_db_column_store_row_group_physical_stats.
object_id
INNER JOIN sys.partitions
ON partitions.object_id = objects.object_id
AND partitions.partition_number = dm_db_column_store_row_group_physical_
stats.partition_number
WHERE objects.name = 'Sale'
AND dm_db_column_store_row_group_physical_stats.has_vertipaq_optimization
IS NOT NULL
AND dm_db_column_store_row_group_physical_stats.has_vertipaq_
optimization = 0
ORDER BY dm_db_column_store_row_group_physical_stats.row_group_id;

The results are shown in Figure 13-6.

Figure 13-6. Rowgroups lacking Vertipaq optimization

217
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

The results show that only three partitions (3, 4, and 5) contain rowgroups that lack
Vertipaq optimization. Therefore, periodic index maintenance could be targeted at
only those partitions, skipping the rest. For less granular details, row_group_id can be
omitted from the query to reduce the result set to a list of partitions that does not contain
rowgroup IDs.

I ndexed Views
It is possible to create indexed views on top of clustered columnstore indexes. While
this represents an added layer of complexity, it allows for queries that do not follow
the natural ordering of the columnstore index to be isolated into their own separate
data structure. There, the view and its indexes can be modified freely without directly
impacting the underlying table.
This can allow for more flexibility when performance testing supplementary indexes
on data that is stored in a clustered columnstore index. The script in Listing 13-9 rebuilds
the columnstore index on Fact.Sale, creates a schemabound view against it, and adds a
pair of indexes to the new view.

Listing 13-9. Script to Rebuild a Table and Add an Indexed View to a Clustered
Columnstore Index
ALTER INDEX [CCX_Fact_Sale] ON fact.sale REBUILD;
GO
CREATE VIEW Fact.v_Sale
WITH SCHEMABINDING
AS
SELECT
    [Sale Key],
    [City Key],
    [Customer Key],
    [Bill To Customer Key],
    [Stock Item Key],
    [Invoice Date Key],
    [Delivery Date Key],
    [Salesperson Key],
    [WWI Invoice ID],

218
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

    Description,
    Package,
    Quantity,
    [Unit Price],
    [Tax Rate],
    [Total Excluding Tax],
    [Tax Amount],
    Profit,
    [Total Including Tax],
    [Total Dry Items],
    [Total Chiller Items],
    [Lineage Key]
FROM Fact.Sale;
GO
CREATE UNIQUE CLUSTERED INDEX CI_v_sale
ON Fact.v_Sale ([Sale Key], [Invoice Date Key]);
GO
CREATE NONCLUSTERED INDEX IX_v_Sale
ON Fact.v_Sale ([Stock Item Key], Quantity)
GO

Rebuilding the columnstore index ensures that the index is in pristine condition
prior to further testing. The view created in this demonstration includes all columns in
the table, but could be revised to include less columns, join other tables, add computed
columns, etc. The next step is to create a unique clustered index on the view. This step
is necessary as nonclustered indexes on a view require a unique index as a prerequisite.
Finally, a nonclustered covering index is added to handle queries against its columns.
Once complete, the analytic query in Listing 13-2 can be modified to target the view
instead, as shown in Listing 13-10.

Listing 13-10. Analytic Query Targeting an Indexed View

SELECT
COUNT(*),
SUM(Quantity) AS total_quantity
FROM Fact.v_Sale
WHERE [Stock Item Key] = 186;

219
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

The results of the query are returned quickly. The execution plan and STATISTICS IO
output are provided in Figure 13-7.

Figure 13-7. Execution plan and STATISTICS IO output for a query against an
indexed view

The execution plan confirms that the nonclustered index on the view v_Sale
was used. The IO output shows acceptably low reads against the view. There is an
unexpected side effect of using an indexed view for managing select reads against a
columnstore index: Vertipaq optimization is still used when new rows are inserted into
the columnstore index!
While this is a useful benefit of indexed views against columnstore indexes, care
should still be exercised when adding complexity to an existing table. Balance the needs
of unusual workloads against a columnstore index vs. the cost of maintaining additional
views and the indexes against those views.
Always test queries without supporting views or indexes first. Confirm their
performance thoroughly and if needed test the impact of additional indexes or indexed
views to confirm whether the additions are worth the cost.

Compression for Nonclustered Rowstore Indexes

Given the inflated size of nonclustered rowstore indexes when added to a clustered
columnstore index, compression can be a valuable way to control the size of additional
indexes and improve read performance against them.

220
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

Consider the nonclustered primary key defined on Fact.Sale. The index size data in
Figure 13-5 shows that this index consumes 6,624KB. The script in Listing 13-11 drops
this constraint and replaces it with a new version that uses page compression.

Listing 13-11. Script to Re-create a Nonclustered Primary Key with Page

Compression
ALTER TABLE Fact.Sale DROP CONSTRAINT PK_Sale;
ALTER TABLE Fact.Sale ADD CONSTRAINT PK_Sale PRIMARY KEY NONCLUSTERED
([Sale Key], [Invoice Date Key])
WITH (DATA_COMPRESSION = PAGE);

When complete, the code in Listing 13-12 can be executed to confirm the new size of
the index, after compression is applied.

Listing 13-12. Script to Check the Size of a Compressed Nonclustered Index

221
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

The results are shown in Figure 13-8.

Figure 13-8. Index size for a page compressed nonclustered primary key

The results show that the index now consumes only 2,736KB, a reduction of 3,888KB
or about 59%. This is a significant space savings and illustrates the positive impact
that compression can have when applied to secondary nonclustered indexes on a
columnstore indexed table.

Nonclustered Rowstore Index Guidance

Adding nonclustered rowstore indexes to a clustered columnstore index is an effective
way of managing frequent, unusual query patterns against analytic data. By splitting
different reads onto both the clustered and nonclustered indexes, contention can be
better managed by also splitting locks onto each index. Before making any indexing
changes, ensure that new indexes are necessary and provide enough long-term benefits
to necessitate their cost.
When possible, consider the use of filters to shrink the size of nonclustered indexes
and decrease their impact on the underlying columnstore index. Similarly, consider
partitioning if not already used as it can assist in streamlining and speeding up index
maintenance processes.
Consider indexed views on top of columnstore indexes when there is added value
in isolating unusual workloads away from the columnstore index and onto a separate
object. The cost of maintaining them is in no way insignificant, but they can provide
a convenient way to encapsulate specific workloads onto their own objects. Splitting
workloads into separate objects helps to reduce contention as reads against each object
target them separately, locking those objects separately from each other.
Regardless of how secondary indexes are created, consider compressing them.
Compression can greatly reduce storage consumption, thereby decreasing memory
footprint and speeding up read queries that use the compressed indexes.

222
Chapter 13 Nonclustered Rowstore Indexes on Columnstore Tables

As a rule of thumb, if the number of nonclustered rowstore indexes against

a columnstore index increases past one (or two with an indexed view), consider
alternative ways of storing this data, such as

• Clustered rowstore table with page compression and supporting

indexes where needed.

• Clustered columnstore index with less nonclustered indexes.

• Separating workloads into separate tables that are managed

independently of each other.

• Apply compression and/or filters to indexes to reduce size and

upkeep of each, making them more manageable.

This chapter provided a variety of tactics toward finding a favorable configuration

that allows wildly different workloads to coexist peacefully on a single table. There is
no one-size-fits-all approach to secondary indexing on a clustered columnstore index,
though. If unsure of the best way to proceed, test each alternative thoroughly to gain
comfort with the pros and cons of each approach.

223
CHAPTER 14

Columnstore Index
Maintenance
Depending on its usage, a columnstore index may require no maintenance at all,
infrequent maintenance, or regular maintenance to ensure optimal storage, resource
consumption, and performance.
Understanding how and when to use index maintenance will ensure that analytic
queries continue to perform well both when a columnstore index is first created and
years into the future.

What Causes Fragmentation?

There are three scenarios that can be used to describe fragmentation in columnstore
indexes:

1. Wasted space due to deleted rows

2. Data that is inserted out of order

3. Multiple delta rowgroups present in the index

When data is deleted in a columnstore index, no rows are actually removed. The
cost to decompress, delete rows, and recompress rowgroups is prohibitively expensive at
runtime. Instead, SQL Server flags the rows as deleted using the delete bitmap, and when
queries access that data, they use the bitmap to identify any rows that are deleted and
need to be skipped.
Chapter 9 provides extensive detail as to how DELETE and UPDATE operations
perform on columnstore indexes and outlines ways of managing them to avoid
persistent performance challenges and minimize fragmentation. The net impact of
deletion on columnstore indexes will be
225
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_14
Chapter 14 Columnstore Index Maintenance

• Space consumed by deleted data.

• More data needs to be read than is needed at runtime.

• Data is spread over more rowgroups than necessary.

The other cause of fragmentation is out-of-order data. A columnstore index should

always be built and maintained with data that is ordered by a key column (or columns)
that are used most frequently in filtering and aggregating queries.
Chapter 10 describes the importance of data order in columnstore indexes and how
maintaining a data order that lines up with common query filters and aggregation can
allow for exceptionally fast query performance.
Data that is inserted regularly in the desired columnstore index order will facilitate
rowgroup elimination, thus allowing vast amounts of the index to be automatically
skipped when filtered using the same column(s). When rows are inserted out of order,
this forces more data to be read than should be necessary. Similarly, since an UPDATE
operation is executed as the combination of a DELETE and INSERT, updating any older
data in a columnstore index will result in it being reinserted at the tail end of the index
alongside newer data.
Consider the example columnstore index in Figure 14-1, which is ordered by
Order_Date.

Figure 14-1. A columnstore index that is ordered by the date dimension

This columnstore index is pristinely ordered by Order Year, with each set of
rowgroups containing rows in ascending order by date. If a query is executed that
calculates metrics for 2016, it can read the first ten rowgroups while ignoring the
other 140.

226
Chapter 14 Columnstore Index Maintenance

Figure 14-2 shows the result of an UPDATE operation that alters data for 100,000
rows in 2016, 50,000 rows in 2017, and 25,000 rows in 2018.

Figure 14-2. Changes in rowgroup structure after an UPDATE operation

When rows were updated, they remained in the columnstore index in their previous
locations, but were flagged as deleted. A new rowgroup was created that contains the
new versions of the deleted rows. The result is that there is now an open rowgroup that
contains data for 2016, 2017, and 2018. In addition, when new data is inserted into the
table, it will also be added to the open rowgroup, resulting in another year’s worth of
data being crammed into a single rowgroup.
Going forward, any query that requires data from 2016, 2017, 2018, or the current
year will need to read this unordered rowgroup. If updates like this are common, then
older rowgroups will quickly become logjammed with deleted rows, while newer
rowgroups become unordered messes of data from many dates. The result will be wasted
storage, wasted memory, and slower queries that need to read far more data than should
be needed to return results.
Unordered inserts will have a similar impact as the INSERT portion of an UPDATE
statement. Inserting data into the columnstore index in Figure 14-1 from 2015 would
result in new rowgroups that contain both new and old data. Unordered inserts will, over
a long period of time, result in the inability of SQL Server to take advantage of rowgroup
elimination as more and more rowgroups contain data from all different time periods.
Delta rowgroups are a key part of columnstore index architecture and ensure
that write operations can occur as quickly as possible. They slow down reads slightly,
though, as rows reside in a b-tree structure and are not compressed with columnstore
compression. The impact of the delta store on read performance is not significant,
but administrators interested in maximizing columnstore read performance would be
interested in compressing them as soon as possible after a data load process completes.

227
Chapter 14 Columnstore Index Maintenance

How Much Fragmentation Is Too Much?

Like with rowstore indexes, fragmentation does not require attention until it becomes
significant enough to impact performance and waste a tangible amount of space. A
columnstore index with 5 billion rows of which 1,000 are deleted should not be seen as
problematic. Alternatively, if one billion of those rows were deleted, then the resulting
fragmentation would be worth addressing.
Fragmentation should be broken out into two distinct calculations, of which each
can be evaluated clinically and acted on appropriately. One is the percentage of a table
that is comprised of deleted rows, and the other is a measure of unordered data within a
columnstore index.

Quantifying Deleted Rows

Deleted rows are included in the view sys.column_store_row_groups and are relatively
easy to report on. The query in Listing 14-1 returns a row per rowgroup within a single
clustered columnstore index.

Listing 14-1. Query to Return Deleted Rows per Rowgroup

SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       partitions.partition_number,
       column_store_row_groups.row_group_id,
       column_store_row_groups.total_rows,
       column_store_row_groups.deleted_rows
FROM sys.column_store_row_groups
INNER JOIN sys.indexes
ON indexes.index_id = column_store_row_groups.index_id
AND indexes.object_id = column_store_row_groups.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.partitions
ON partitions.partition_number = column_store_row_groups.partition_number
AND partitions.index_id = indexes.index_id

228
Chapter 14 Columnstore Index Maintenance

AND partitions.object_id = tables.object_id

WHERE tables.name = 'Sale_CCI'
ORDER BY indexes.index_id, column_store_row_groups.row_group_id;

The results are shown in Figure 14-3.

Figure 14-3. Deleted row counts per rowgroup for a columnstore index

This detail shows how many rows are deleted per rowgroup. The results can
be aggregated to show deleted rows per partition or for the entire table. If a table is
partitioned, then knowing if the deleted rows exist only in one partition or all of them is
useful for determining if all or only some partitions require attention.
In the example results outlined in Figure 14-3, about 5% of the columnstore index is
comprised of deleted rows that are spread across the index relatively evenly. As a rough
guideline, there is no urgent need for index maintenance to address this unless the
percentage of deleted rows is at least 10% of the total rows in a partition or in the table.
Keep in mind that the performance impact of deleted rows is gradual over time.
There will never be a scenario where a threshold is reached in which performance
suddenly plummets. Therefore, automating index maintenance to occur when deleted
rows exceed some set percentage is an effective way to avoid accidentally allowing an
index to become absurdly fragmented.

229
Chapter 14 Columnstore Index Maintenance

Detailing Unordered Data

Understanding the effectiveness of ordered data is easy to see via querying segment
metadata, but can be challenging to quantify. The query in Listing 14-2 shows how to
retrieve the segment minimum and maximum data IDs for a given columnstore index
and a single column.

Listing 14-2. Query to Retrieve Min/Max Data IDs for a Given Column in a
Columnstore Index
SELECT
       tables.name AS table_name,
       indexes.name AS index_name,
       columns.name AS column_name,
       partitions.partition_number,
       column_store_segments.segment_id,
       column_store_segments.min_data_id,
       column_store_segments.max_data_id,
       column_store_segments.row_count
FROM sys.column_store_segments
INNER JOIN sys.partitions
ON column_store_segments.hobt_id = partitions.hobt_id
INNER JOIN sys.indexes
ON indexes.index_id = partitions.index_id
AND indexes.object_id = partitions.object_id
INNER JOIN sys.tables
ON tables.object_id = indexes.object_id
INNER JOIN sys.columns
ON tables.object_id = columns.object_id
AND column_store_segments.column_id = columns.column_id
WHERE tables.name = 'Sale_CCI'
AND columns.name = 'Invoice Date Key'
ORDER BY tables.name, columns.name, column_store_segments.segment_id;

The results of this query provide useful insight into the data order within this index,
as shown in Figure 14-4.

230
Chapter 14 Columnstore Index Maintenance

Figure 14-4. Segment minimum/maximum value metadata for an unordered

columnstore index

In the metadata, note that the values for min_data_id and max_data_id are the same
for each rowgroup. This means that queries that filter on the column Invoice Date Key
will be forced to scan the entire table to return results as any value could be found in any
rowgroup.
Consider an ordered version of this table, where the results are as seen in Figure 14-5.

Figure 14-5. Segment minimum/maximum value metadata for an ordered

columnstore index

This version of the table shows a progression of values for min_data_id and
max_data_id that increases as the segment_id increases. Because each segment contains
a distinct grouping of column values, this metadata can be effectively used to skip any
rowgroups that contain values that are irrelevant to a query. The query in Listing 14-3
returns a complete list of all segments in a columnstore index that have any overlapping
values. The inequalities are not inclusive as it is common that the start and end values in
a rowgroup will overlap those in the next rowgroup.

231
Chapter 14 Columnstore Index Maintenance

Listing 14-3. Query to Return a List of Overlapping Segments

WITH CTE_SEGMENTS AS (
       SELECT
              tables.name AS table_name,
              indexes.name AS index_name,
              columns.name AS column_name,
              partitions.partition_number,
              column_store_segments.segment_id,
              column_store_segments.min_data_id,
              column_store_segments.max_data_id,
              column_store_segments.row_count
       FROM sys.column_store_segments
       INNER JOIN sys.partitions
       ON column_store_segments.hobt_id = partitions.hobt_id
       INNER JOIN sys.indexes
       ON indexes.index_id = partitions.index_id
       AND indexes.object_id = partitions.object_id
       INNER JOIN sys.tables
       ON tables.object_id = indexes.object_id
       INNER JOIN sys.columns
       ON tables.object_id = columns.object_id
       AND column_store_segments.column_id = columns.column_id
       WHERE tables.name = 'Sale_CCI_ORDERED'
       AND columns.name = 'Invoice Date Key')
SELECT
       CTE_SEGMENTS.table_name,
       CTE_SEGMENTS.index_name,
       CTE_SEGMENTS.column_name,
       CTE_SEGMENTS.partition_number,
       CTE_SEGMENTS.segment_id,
       CTE_SEGMENTS.min_data_id,
       CTE_SEGMENTS.max_data_id,
       CTE_SEGMENTS.row_count,
       OVERLAPPING_SEGMENT.partition_number AS
overlapping_partition_number,

232
Chapter 14 Columnstore Index Maintenance

       OVERLAPPING_SEGMENT.segment_id AS overlapping_segment_id,
       OVERLAPPING_SEGMENT.min_data_id AS overlapping_min_data_id,
       OVERLAPPING_SEGMENT.max_data_id AS overlapping_max_data_id
FROM CTE_SEGMENTS
INNER JOIN CTE_SEGMENTS OVERLAPPING_SEGMENT
ON (OVERLAPPING_SEGMENT.min_data_id > CTE_SEGMENTS.min_data_id
AND OVERLAPPING_SEGMENT.min_data_id < CTE_SEGMENTS.max_data_id)
OR (OVERLAPPING_SEGMENT.max_data_id > CTE_SEGMENTS.min_data_id
AND OVERLAPPING_SEGMENT.max_data_id < CTE_SEGMENTS.max_data_id)
OR (OVERLAPPING_SEGMENT.min_data_id < CTE_SEGMENTS.min_data_id
AND OVERLAPPING_SEGMENT.max_data_id > CTE_SEGMENTS.max_data_id)
ORDER BY CTE_SEGMENTS.partition_number, CTE_SEGMENTS.segment_id

This query evaluates the boundaries for the minimum and maximum value of one
column within each rowgroup and determines if any other rowgroups overlap those values.
The list returned by the query in Figure 14-6 may appear long at first glance, but it
is important to note that any columnstore index that has been the target of UPDATE
operations or unordered inserts will have entries here. Looking at the data returned, it
can be seen that the out-of-order data is spread somewhat evenly across segments, with
each segment containing 1–4 other segments that overlap at least one value with it.

Figure 14-6. List of overlapping values within rowgroups for the Invoice Date
Key column

While there is no precise way to measure the percentage of unordered data in the
same way that it was possible to measure the percentage of rows in a columnstore
index that are deleted, it is possible to gauge how effectively data order impacts query
performance by performing metadata tests using COUNT(*) queries against the
columnstore index. This could be done for every date in the table, which would result in
a very thorough experiment. For the sake of a simple demonstration, eight sample dates
will be chosen at random to test, as given by the query in Listing 14-4.

233
Chapter 14 Columnstore Index Maintenance

Listing 14-4. Sample Dates to Test How Ordered Data Is in a Columnstore Index

SELECT
[Invoice Date Key],
COUNT(*) AS Sale_Count
FROM Fact.Sale_CCI_ORDERED
WHERE [Invoice Date Key] IN ('5/1/2013', '9/5/2013', '1/17/2014',
'6/30/2014', '3/14/2015', '12/12/2015', '1/1/2016', '2/29/2016')
GROUP BY [Invoice Date Key]
ORDER BY [Invoice Date Key]

The results in Figure 14-7 provide row counts for each data.

Figure 14-7. List of sample dates for use in testing the effectiveness of columnstore
data order

Each data point chosen contains at most 0.1% of the data in the table as there are
25,109,150 total rows, of which these encompass eight of its dates. Based on the row
counts and table size, an ideal ordered table would only require reading 1–2 rowgroups
to retrieve data for any of those given dates. The query in Listing 14-5 executes separate
COUNT(*) queries for each date identified earlier.

Listing 14-5. COUNT Queries to Test Rowgroup Reads

SET NOCOUNT ON;

SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'5/1/2013';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'9/5/2013';

234
Chapter 14 Columnstore Index Maintenance

SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =

'1/17/2014';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'6/30/2014';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'3/14/2015';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'12/12/2015';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'1/1/2016';
SELECT COUNT(*) FROM Fact.Sale_CCI_ORDERED WHERE [Invoice Date Key] =
'2/29/2016';

Figure 14-8 shows the STATISTICS IO output for each preceding query.

Figure 14-8. STATISTICS IO output for COUNT queries in Listing 14-5

For each of the eight COUNT(*) sample queries, 3–4 rowgroups are read in order
to retrieve the count. Based on the knowledge that these queries should read 1–2
rowgroups, it can be deduced that queries are generally reading two to three times more
rowgroups than is necessary. Depending on how often updates and unordered inserts

235
Chapter 14 Columnstore Index Maintenance

occur on the table, this may be acceptable or it may be unusual. Knowledge of the table’s
usage helps in understanding how extreme these numbers are and if reading three
rowgroups instead of one is normal or worthy of attention.
Testing rowgroups read and skipped using STATISTICS IO is an effective way to
gauge how ordered a columnstore index is. If there is a desire to be complete about this
test and use count queries against many (or all) dates in the table, consider treating them
as maintenance and performing that research at a predetermined time when running a
lengthy maintenance script is acceptable.

The No-Maintenance Columnstore Index

An ideal columnstore index rarely requires maintenance. This is possible when the
following are true:

• Rows are never deleted.

• Rows are never updated.

• Rows are always inserted in order.

If those three conditions can be met, then a columnstore index can be spared of
nearly all maintenance. Whether any is performed is up to the whim of its administrator.
Realistically, the only suboptimal scenario that can arise when the columnstore index
is not the target of deletes, updates, or unordered inserts is that rowgroups may be
undersized due to the use of the delta store to process small INSERT operations.
The impact of undersized rowgroups resulting from delta store usage is minor and
should not be seen as an urgent problem in need of an immediate resolution. In these
scenarios, waiting for infrequent maintenance periods to perform columnstore index
maintenance would be more than effective enough to ensure that undersized rowgroups
are merged effectively. Quarterly or yearly is likely often enough.
Now that the causes of fragmentation have been thoroughly detailed, using index
maintenance operations to resolve fragmentation can be discussed.

236
Chapter 14 Columnstore Index Maintenance

Columnstore Reorganize
The fastest and simplest operation available for a columnstore index is the
REORGANIZE. This is an online operation that accomplishes the following tasks:

• Combines undersized rowgroups via a columnstore index merge.

This occurs when multiple rowgroups can combine to fit into a single
new rowgroup (less than 1,024,576 or 220 rows).
• May remove deleted rows via a self-merge. This only happens when
more than 102,400 rows in a rowgroup are deleted.

If both of these tasks apply to a set of rowgroups, then a merge operation will be
prioritized over a self-merge. Rowgroups that were trimmed due to dictionary pressure
cannot be combined with other rowgroups, regardless of their row counts.
To demonstrate the merge and self-merge operations that can be used by a
columnstore REORGANIZE operation, a large set of rows will be deleted from a
columnstore index, as seen in Listing 14-6

Listing 14-6. Query to Delete a Large Portion of a Rowgroup

DELETE
FROM Fact.Sale_CCI_ORDERED
WHERE [Invoice Date Key] <= '1/17/2013';

Figure 14-9 shows the resulting set of deleted rows within the columnstore index’s
rowgroup metadata using the same query as provided in Listing 14-1.

Figure 14-9. Deleted rows per rowgroup for a columnstore index

237
Chapter 14 Columnstore Index Maintenance

The second column to the right shows total rows in the rowgroup, whereas the
rightmost column provides the deleted row count. Of the rowgroups with deleted rows,
only rowgroups 23 and 24 have more than 102,400 rows deleted and would qualify for a
self-merge operation. These rowgroups are also valid targets for a columnstore merge
operation as they can be combined, with the resulting rowgroup containing less than the
row cap (220 rows) for a columnstore rowgroup.
The syntax for a columnstore REORGANIZE operation is shown in Listing 14-7.

Listing 14-7. Syntax for a Columnstore Reorganize Operation

ALTER INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED REORGANIZE;

After the REORGANIZE operation is complete, the rowgroup metadata can be

reviewed again, as shown in Figure 14-10.

Figure 14-10. Rowgroup metadata following an index REORGANIZE operation

Note that rowgroups 23 and 24 are now flagged as TOMBSTONE and will be cleaned
up by the tuple mover at some point in the near future. Two new rowgroups were created
(25 and 26) that replace them, with the deleted rows removed. The self-merge operation
essentially creates new rowgroups, copies all nondeleted rows into them, and swaps
them in as the active rowgroups while the previous versions are flagged for cleanup.
The resulting rowgroups are free of the burden of deleted rows. Remember that the
self-merge only occurs when more than 102,400 rows in a rowgroup are deleted.
After a minute passes, the rowgroup metadata confirms that the rowgroups labeled
as TOMBSTONE are now removed from the columnstore index, as seen in Figure 14-11.
This is an automatic cleanup process that requires no operator intervention to trigger.

238
Chapter 14 Columnstore Index Maintenance

Figure 14-11. Rowgroup metadata after the tuple mover removes TOMBSTONE
rowgroups

Rowgroups can only be combined if the reason they are undersized is not related to
dictionary pressure. Figure 14-12 shows additional metadata from sys.dm_db_column_
store_row_group_physical_stats for this columnstore index.

Figure 14-12. Rowgroup metadata for a columnstore index, including rowgroup

trim reason

Because rowgroup 25 was trimmed due to dictionary pressure, it cannot be

combined with rowgroup 26, even though their combined row counts could fit them into
a single rowgroup.

Reorganize to Remove Delta Rowgroups

An additional option is available when issuing a columnstore index REORGANIZE
command: the COMPRESS_ALL_ROW_GROUPS option. When used, SQL Server will
initialize the tuple mover to process the contents of the delta store and move them into
compressed rowgroups. This effectively empties the delta store, moving all data into
compressed rowgroups.

239
Chapter 14 Columnstore Index Maintenance

The upside of this operation is that it ensures faster read operations, as there is no
need to read the rowstore structure of the delta store when processing queries. The
downside is that it is an additional maintenance option that requires time and resources
to execute. Consider the INSERT operation shown in Listing 14-8.

Listing 14-8. Small INSERT Operation into a Columnstore Index

INSERT INTO Fact.Sale_CCI_ORDERED

One row is inserted into the columnstore index. Rowgroup metadata can confirm the
new row that resides in an open delta rowgroup, as seen in Figure 14-13, using the same
query as in Listing 14-1.

Figure 14-13. Rowgroup metadata for a newly inserted row into a delta rowgroup

With a single row in the delta store, a REORGANIZE operation will be run using the
COMPRESS_ALL_ROW_GROUPS option, as seen in Listing 14-9.

Listing 14-9. REORGANIZE Using the COMPRESS_ALL_ROW_GROUPS option

ALTER INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED REORGANIZE

WITH (COMPRESS_ALL_ROW_GROUPS = ON);

240
Chapter 14 Columnstore Index Maintenance

The resulting columnstore metadata after the index maintenance is shown in

Figure 14-14.

Figure 14-14. Rowgroup metadata immediately after index maintenance

The metadata shows a single row in a compressed rowgroup, and the old delta store
set to TOMBSTONE, awaiting cleanup. Figure 14-15 shows the rowgroup metadata after
the TOMBSTONE rowgroup is cleaned up.

Figure 14-15. Rowgroup metadata after garbage collection

A compressed rowgroup of a single row is silly, so another REORGANIZE will

be executed against the columnstore index. The resulting metadata can be seen in
Figure 14-16.

Figure 14-16. Rowgroup metadata after undersized rowgroups are combined

Rowgroup 25 remains undersized due to dictionary pressure, but rowgroups 26 and

28 have been combined into the newly formed rowgroup 29. The trim reason description
allows an operator to understand that this rowgroup was created and populated via an
index REORGANIZE operation.

241
Chapter 14 Columnstore Index Maintenance

Index REORGANIZE operations are an excellent way to combine undersized

rowgroups, assuming they do not suffer from dictionary size limitations. They can also
remove deleted rows, assuming the number of deleted rows is greater than 102,400.
Lastly, with the COMPRESS_ALL_ROW_GROUPS option, an index REORGANIZE
can process the delta store contents into compressed rowgroups, which can later be
combined via further REORGANIZE operations, if desired.
For a columnstore index that is not the target of UPDATE operations and does not
experience out-of-order inserts, using index REORGANIZE operations will handle most
fragmentation effectively and preclude the need for other regular index maintenance
tasks. The simplest application of index maintenance is to run an index REORGANIZE
operation after each data load and include the COMPRESS_ALL_ROW_GROUPS option.
This will ensure that rowgroups remain relatively full and that delta stores are processed
and not slowing down columnstore queries. For a stickler that wants perfection, a
second REORGANIZE operation can be executed to further combine rowgroups
resulting from processed delta stores. These are online, fast, and effective ways to
manage fragmentation on an ongoing basis as a columnstore index grows via regular
data load processes.

Columnstore Rebuild
Rebuilding a columnstore index functions similarly to rebuilding a rowstore index.
When a REBUILD is issued, a completely new copy of the columnstore index is created,
replacing the old index. While an expensive process, the results are
• Rowgroups that are filled to capacity, whenever possible.

• All deleted rows are eliminated.

• All delta stores are processed and compressed.

• Vertipaq optimization is applied, even if it previously was not.

In essence, a REBUILD is roughly equivalent to creating a columnstore index

anew using a CREATE INDEX statement. The only difference is that the existing index
structure will remain intact until the new index is built. For ONLINE rebuilds, this

242
Chapter 14 Columnstore Index Maintenance

ensures that queries can continue to use the columnstore index, even as a rebuild
operation is running.
Rebuilding a clustered columnstore index can only be accomplished as an online
operation starting in SQL Server 2019. Nonclustered indexes can be rebuilt online
regardless of SQL Server version.
Consider the columnstore index that has been tested recently in this chapter. The
T-SQL in Listing 14-10 issues a REBUILD against that index.

Listing 14-10. REBUILD Statement Against a Clustered Columnstore Index

ALTER INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED REBUILD;

A REBUILD operation takes significantly longer than a REORGANIZE as the entire

index must be re-created in its entirety. Alternatively, a REORGANIZE follows specific
rules to determine what work to do and will usually only operate on a small portion of a
columnstore index. Figure 14-17 shows a columnstore index after a REBUILD operation
completes.

Figure 14-17. Rowgroup metadata after a columnstore index REBUILD

After a REBUILD, the columnstore index has no deleted rows, and rowgroups
are mostly full. Oddly enough, the last two rowgroups (23 and 24) are undersized
residuals and can be cleaned up via an index REORGANIZE operation. The metadata
in Figure 14-18 shows the final results after the index is subject to an additional
REORGANIZE after the REBUILD.

243
Chapter 14 Columnstore Index Maintenance

Figure 14-18. Rowgroup metadata after a columnstore index REBUILD and

REORGANIZE

Finally, the index is pristine, with 23 rowgroups that are completely full and 1
additional rowgroup that is leftover from the index maintenance operations.
Index REBUILD operations should be used infrequently to manage one of a few
scenarios:

• Undersized rowgroups accumulated over a long period of time that

cannot be combined via REORGANIZE operations.

• Extensive deleted rows that number less than 102,400 rows per
rowgroup and cannot be addressed by REORGANIZE operations.

• Large amounts of out-of-order data exist in the columnstore index.

This must be addressed in conjunction with work to reorder data, as
shown in the next section.

Note that when an index REBUILD is issued, the compression type may be changed
to or from columnstore archive compression, if needed. Because an index REBUILD
is an expensive operation and is offline prior to SQL Server 2019, care should be taken
to execute rebuilds during maintenance windows when such work is more tolerable to
processes that use this data.
Sometimes, data can become heavily fragmented via software releases that make
significant changes to the underlying schema and data. If this is anticipated, then
scheduling an index REBUILD as a postprocessing step after the software release would
be an excellent way to ensure that data continues to be efficiently delivered to analytic
processes, even after a disruptive software release.
REBUILD operations may target a specific partition, thus allowing only data that is
heavily fragmented to be rebuilt. For a large columnstore index in which only a small
portion is actively written to, this is an excellent way to speed up rebuilds and minimize
disruptions to the users of analytic data.

244
Chapter 14 Columnstore Index Maintenance

Columnstore Reorder and Rebuild

Clustered columnstore indexes do not enforce data order. It is up to the data architect
that designs these tables to determine how to order data and to ensure that order is
enforced via any processes that write to the table. The only way to effectively resolve the
challenge of unordered columnstore data (without building new tables and structures) is
to perform the following tasks:

1. Drop the columnstore index.

2. Create a clustered rowstore index with key columns that match

the data order for the table.

3. Create a new clustered columnstore index with the

DROP_EXISTING=ON option.

The clustered rowstore index is used to enforce a new data order on the contents of
the table, whereas the new columnstore index replaces it, retaining the new data order.
This is an expensive and disruptive process that will result in analytic queries
being unable to take advantage of the columnstore index from the point when the
index is dropped until the new columnstore index is completely built. Therefore, it is a
worthwhile process to implement during a maintenance window when causing trouble
is more acceptable.
The script in Listing 14-11 performs these operations to reorder data from its current
state to data that is ordered perfectly by Invoice Date Key.

Listing 14-11. Process to Fix Poorly Ordered Data in a Clustered

Columnstore Index
/*     Step 1 */
DROP INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED;
/*     Step 2 */
CREATE CLUSTERED INDEX CCI_fact_Sale_CCI_ORDERED ON Fact.Sale_CCI_ORDERED
([Invoice Date Key]);
/*     Step 3 */
CREATE CLUSTERED COLUMNSTORE INDEX CCI_fact_Sale_CCI_ORDERED ON
Fact.Sale_CCI_ORDERED WITH (DROP_EXISTING=ON, MAXDOP=1);

245
Chapter 14 Columnstore Index Maintenance

Note that dropping a clustered columnstore index is not a speedy operation. It

requires converting the columnstore structure into a b-tree/heap structure, which
requires more computing resources than dropping a clustered rowstore index.
The process to fix data order on a columnstore index should be reserved for
scenarios where fragmentation is high enough to negatively impact analytic processes
and should be an infrequent operation.
If data becomes unordered quickly in a columnstore index, then consider ways to
alter how data load processes operate to reduce fragmentation via those data loads. The
following are a few helpful guidelines for doing so:

• Remove all UPDATE operations on columnstore indexes. Perform

updates on staging tables instead (when possible).

• Avoid inserting data out of order. Use temporary or staging tables to

ensure that data is ordered correctly prior to inserting new data.

Despite its complexity, the process to fix unordered data can be executed on a
partition-by-partition basis, thus allowing only active partitions containing warm/hot
data to be targeted with expensive index maintenance operations. For a columnstore
index with many partitions, this can save considerable time and reduce disruption and
downtime to analytic processes.

Columnstore Index Maintenance by Partition

The syntax to rebuild a specific partition of a columnstore index is shown in
Listing 14-12.

Listing 14-12. Query to Rebuild a Single Partition of a Columnstore Index on a

Partitioned Table
ALTER INDEX CCI_fact_Sale_CCI_PARTITIONED ON Fact.Sale_CCI_PARTITIONED
REBUILD PARTITION = 5;

In this example, if partition number 5 is identified as the only partition containing

data that is being loaded/modified regularly, then rebuilding only it will save significant
maintenance time as the other partitions can be skipped.

246
Chapter 14 Columnstore Index Maintenance

Index Maintenance in Nonclustered

Columnstore Indexes
Nonclustered columnstore indexes are afforded less flexibility with regard to index
maintenance. Data order is enforced by the clustered rowstore index, and therefore
the quality of data order within a columnstore index is prescribed by its corresponding
clustered rowstore index. If the table’s clustered rowstore index happens to be the
ordering column(s) for the columnstore index and that table is not subject to UPDATE
operations, then the nonclustered columnstore index will maintain data order effectively.
Otherwise, index maintenance for nonclustered columnstore indexes will operate
similarly to maintenance on clustered columnstore indexes:

• Use REORGANIZE operations to combine undersized rowgroups,

process the delta store, and eliminate larger groups of deleted rows.

• Use REBUILD operations to eliminate undersized rowgroups and

eliminate unordered data due to UPDATE operations.

The primary difference here is that a REBUILD can eliminate out-of-order data
resulting from UPDATE operations. When rebuilt, a nonclustered columnstore index
will be created using the order prescribed by the clustered rowstore index, which is not
negatively impacted by UPDATE operations in the same way as a columnstore index.
Both REORGANIZE and REBUILD operations are available as online operations
for nonclustered indexes, providing more flexibility when trying to schedule recurring
(or one-time) maintenance. This means that real-time operational analytics that target
a nonclustered columnstore index can continue efficiently, even as that index is being
rebuilt.
Like with clustered columnstore indexes, maintenance on nonclustered columnstore
indexes can be issued against any or all partitions, allowing active data to be targeted
with maintenance while unchanging data can be skipped.

247
CHAPTER 15

Columnstore Index
Performance
The ultimate measure of performance for any data structure is the speed in which data
can be retrieved. In columnstore indexes, the time required to return data will be a
function of two operations:

• Metadata reads

• Data reads

This chapter will walk through the steps needed to measure, assess, and tune
columnstore index query performance, both for analytics and write operations. It
will also introduce additional options for columnstore index usage and performance
optimization.

Columnstore Metadata Reads

To determine what data to read and how to read it, the metadata for a columnstore index
will be read prior to any columnstore index scans. This includes the metadata about the
contents of rowgroups and segments. Consider the query in Listing 15-1.

Listing 15-1. Example Query Used in Performance Analysis

SELECT
SUM(Quantity) AS Total_Quantity,
SUM([Total Excluding Tax]) AS Total_Excluding_Tax
FROM Fact.Sale_CCI_PARTITIONED
WHERE [Invoice Date Key] = '7/17/2015';

249
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5_15
Chapter 15 Columnstore Index Performance

Two columns are aggregated for a single value of Invoice Date Key. This table is
both ordered and partitioned and therefore will benefit significantly from partition
elimination and rowgroup elimination. When the query is executed, columnstore
metadata is consulted to determine which segments need to be read.
The table Fact.Sale_CCI_PARTITIONED is partitioned on the Invoice Date Key
column by year, with partitions assigned to the years 2013, 2014, 2015, 2016, and 2017.
Listing 15-2 provides the definitions for both the partition scheme and function used in
this demonstration.

Listing 15-2. Partition Function and Partition Scheme Definitions

CREATE PARTITION FUNCTION fact_Sale_CCI_years_function (DATE) AS RANGE

RIGHT FOR VALUES
('1/1/2014', '1/1/2015', '1/1/2016', '1/1/2017');

CREATE PARTITION SCHEME fact_Sale_CCI_years_scheme AS PARTITION

fact_Sale_CCI_years_function
TO (WideWorldImportersDW_2013_fg, WideWorldImportersDW_2014_fg,
WideWorldImportersDW_2015_fg,
WideWorldImportersDW_2016_fg, WideWorldImportersDW_2017_fg)

The table contains data ranging from 2013 through 2016 and therefore will not make
use of the last partition. When executed, the query in Listing 15-1 will immediately check
the filter against the partition function and then use the partition scheme to determine
where data is located based on the function. Based on this information, it is determined
that only data in the partition WideWorldImportersDW_2015_fg will be needed.
Columnstore metadata is stored separately for each partition, and therefore when
executed, this query needs to only consult metadata relevant to the partition containing
data for the year 2015. Figure 15-1 shows the segment metadata for Invoice Date Key in
this columnstore index.

250
Chapter 15 Columnstore Index Performance

Figure 15-1. Columnstore segment metadata for Invoice Date Key

The metadata shows an exceptionally well-ordered columnstore index. Note that

min_data_id and max_data_id progressively increase as values advance through each
segment within each partition. An encoding type of 2 represents dictionary encoding for
a nonstring data type, in this case a DATE. The primary dictionary used for this column
has an ID of zero.
The query in Listing 15-3 provides a bit of additional detail on the dictionary used for
this column.

Listing 15-3. Query to Return Information About a Columnstore Dictionary

SELECT
       partitions.partition_number,
       objects.name,
       columns.name,
       column_store_dictionaries.type,
       column_store_dictionaries.entry_count,
       column_store_dictionaries.on_disk_size
FROM sys.column_store_dictionaries
INNER JOIN sys.partitions
ON column_store_dictionaries.hobt_id = partitions.hobt_id
INNER JOIN sys.objects
ON objects.object_id = partitions.object_id
INNER JOIN sys.columns
ON columns.column_id = column_store_dictionaries.column_id
AND columns.object_id = objects.object_id
WHERE objects.name = 'Sale_CCI_PARTITIONED'
AND columns.name = 'Invoice Date Key'
AND column_store_dictionaries.dictionary_id = 0;
251
Chapter 15 Columnstore Index Performance

The results provide additional detail about the dictionary used for this column, as
shown in Figure 15-2.

Figure 15-2. Columnstore dictionary metadata

This shows that partitions 1–3 share a single dictionary that is a compact 1308 bytes.
The scenario presented here is essentially the optimal columnstore index and query. The
index is well ordered and partitioned and the query aligns perfectly with that order to
minimize the amount of metadata read as part of its execution.
While columnstore metadata may not seem large when compared to the data itself,
it is important to remember that a columnstore index with one billion rows will have
approximately one thousand rowgroups. Each of those rowgroups will contain one
segment per column. For this table which contains 21 columns, metadata will consist
of about 21,000 segments per billion rows. Needing to read metadata on rowgroups,
segments, dictionaries, and more can add up to a nontrivial amount of work for SQL
Server. As with data itself, metadata needs to be loaded into memory before it can be
read. Therefore, maintaining an ordered table and partitioning (if needed) can ensure
that excessive metadata is not read when executing analytic queries.
Returning to the analytic query in Listing 15-1, it can be expected that it would
execute quickly, resulting in minimal metadata reads, as well as minimal data reads. The
STATISTICS IO output for the query is shown in Figure 15-3.

Figure 15-3. Statistics IO output for a sample analytic query

252
Chapter 15 Columnstore Index Performance

This example was an ideal scenario where segment elimination, rowgroup

elimination, and partitioning all combined to provide exceptionally low reads and a vast
majority of segments skipped. While the real world will not always provide such optimal
examples of columnstore performance, this illustrates the upper bound for what to
compare performance against.

Columnstore Data Reads

Once columnstore metadata is read, the data itself is then read. It is critical to
understand that the performance of queries against a columnstore index will ultimately
be driven by the number of segments read when a query is executed. All other
variables being equal, an analytic request that reads 1000 segments will be expected to
perform approximately ten times slower than a query that reads 100 segments. While
columnstore indexes may at times be seen as magical tools sprinkled with pixie dust, this
is the one area in which there is no magic and performance will be guided by IO.
Segment reads equate to page reads which represent those IO operations. If the
pages needed are already in memory, then the operations will be comprised of logical
IO as segments are read directly from the buffer pool. If the required pages are not in
memory yet (or are partially in memory), then physical reads will be required to copy
pages from storage into the buffer pool.
Segment elimination and rowgroup elimination are the tools that limit segment
reads and help ensure that only the amount of data needed to satisfy a query is read. The
quality of columnstore compression applied to data helps squeeze data into as few pages
as possible, thereby reducing IO.
At their most basic structure, columnstore indexes are bound by the same speed
limitations that exist for rowstore tables. Reading pages from storage into the buffer pool
takes time. Decompressing and reading data from the buffer pool takes additional time.
Therefore, the less pages that need to be read, the faster columnstore index analytics can
perform.
Other performance-boosting tools used by columnstore indexes (such as batch
mode operation or filtered indexes) further improve query speeds by improving
throughput or reducing IO. The focus throughout the step of reading columnstore data is
to control IO via whatever means possible. Figure 15-4 illustrates how data and metadata
are read from storage through to being used query results.

253
Chapter 15 Columnstore Index Performance

Figure 15-4. Movement of columnstore data and metadata from storage to

memory to usage by a query

While the entire columnstore index resides on a storage system, typically only
a fraction of it will be maintained in the buffer pool. Segments remain compressed
throughout this entire IO process until they are needed by a query, at which time they
are decompressed, and data is returned to SQL Server.

Memory Sizing
All of the performance discussions thus far have focused on reducing IO. Despite all
efforts to minimize unnecessary IO, larger columnstore indexes will ultimately require
reading significant numbers of segments from storage into memory.
The next step in ensuring highly performant columnstore indexes is to accurately
estimate the amount of memory that should be allocated to SQL Server to support
analytic queries.
Too little memory and data will be constantly removed from memory, only to be
read back into the buffer pool when needed again soon. The process of reading data
from storage (even fast storage) will be far greater than other steps in executing analytic
queries.
Too much memory would represent unused computing resources, which translate
into wasted money.

254
Chapter 15 Columnstore Index Performance

Allocating memory to SQL Server to support columnstore indexes requires a handful

of steps:

1. Measure the size of the columnstore index.

2. Approximate how much of the index is comprised of hot data that

should always reside in memory.

3. Approximate how much of the index represents warm data. This

data will be in memory for some analytic needs, but not all.

4. Approximate data growth over time that will adjust the numbers
calculated earlier.

Any data that is not hot or warm is expected to be cold and rarely used. It may still
be valuable but is not accessed often enough to architect an entire system around it.
Resources can be allocated at runtime by Azure, AWS, or other hosting services, if that
level of flexibility is desired, but that is entirely optional and at the discretion of an
organization and the importance of speedy access to cold data.
Consider a hypothetical columnstore index as shown in Listing 15-4.

Listing 15-4. Row Counts and Size for Data in a Columnstore Index

2017: 10,000,000 rows (1GB) - cold/rarely used

2018: 50,000,000 rows (3GB) - cold/rarely used
2019: 100,000,000 rows (5GB) - cold/rarely used
2020: 250,000,000 rows (10GB) - warm/sometimes used
2021: 500,000,000 rows (18GB) - hot/often used

This index shows a common trend in analytics by which the most recent year is
heavily used for current reporting. The previous year is also accessed frequently for year-
over-year analytics. All previous data is important and must be retained, but is not read
very often when compared to newer data.
Based on the numbers provided, if the goal was to ensure that the most frequently
used data was accounted for in memory allocation, then this columnstore index would
warrant 18GB of memory to ensure that all 18GB of hot data can reside in memory,
if needed.
If the 10GB of warm data were also important, then allocating some amount of
memory up to 10GB would help ensure that data is not cycled in and out of memory too
often and that it does not replace more important hot data in memory when requested.

255
Chapter 15 Columnstore Index Performance

If resources were plentiful, then allocating the full 10GB would accomplish the task in its
entirety. Otherwise, the organization responsible for this data would need to determine
how important the warm data is and allocate some amount of additional memory up
to 10GB to cover it. For example, if it was estimated that the latter half of the 2020 data
would be requested on a somewhat regular basis, but the earlier half would be far
less used, then 5GB would be a good estimate of memory to allocate to this block of
warm data.
The remaining 9GB of data from 2017 to 2019 would not receive any memory
allocation as they are rarely read and would impact performance too infrequently to
matter to most organizations. If infrequent analytics or reports are disruptive enough, an
argument could be made to adding further memory to reduce that disruption, but this
would be something that should be considered on a case-by-case basis.
This example also illustrates a fairly straightforward growth of about two times per
year. If this rate of growth is expected to continue and the years with warm and hot data
are expected to roll forward each year, then in a year, the current hot data (18GB) would
become warm data, the current warm data (10GB) would become cold data, and next
year’s data (~35GB) would become the new hot data. Therefore, year-over-year growth
would be represented by

(35GB + 18GB) - (18GB + 10GB) = 25GB.

Note that the size of columnstore data used in memory will often not be the same
as the total data size. A columnstore index with 20 columns may not see all columns
used equally. It is quite possible that some columns may rarely get used. If this is the
case, then memory estimations can adjust the amount needed downward to account
for columns that are not needed in memory often. While these infrequently queried
columns cannot be classified as cold data, they can be discounted from memory totals to
ensure that memory estimates are not overinflated by segments that are rarely needed.
For example, if the 18GB of data for 2021 included 5GB of columns that are rarely
queried, then the memory estimate for that data could be reduced from 18GB to as little
as 13GB, if there is confidence in how infrequently those columns are used.
There is a great deal of estimation that goes into this work. At the absolute high end,
a memory budget could encompass all space consumed by a columnstore index. At the
lowest level, no memory could be allocated for these analytics. A realistic estimation will
be somewhere in the middle and will be influenced by the performance required of this
data and the availability/cost of computing resources.

256
Chapter 15 Columnstore Index Performance

Dictionary Size and Dictionary Pressure

Dictionary encoding is one of a multitude of ways in which SQL Server can improve
columnstore compression effectiveness. Chapter 5 discussed encoding methods in detail
and how dictionary encoding can be used by columnstore indexes to greatly reduce
index size for columns that have repetitive values.
While dictionaries can be shared across many segments, a caveat to dictionary
encoding is that they cap at 16 megabytes. As a result, if a column has wide values that do
not repeat often, then the result may be that a dictionary fills up before a rowgroup could
be completely written. The scenario when a dictionary fills up is known as dictionary
pressure. While a dictionary is limited to 16MB, dictionary pressure will often occur at
sizes well below that as the limitation is on how much space will be consumed there,
rather than how much is currently used.
When a columnstore index experiences dictionary pressure, the rowgroup that is
currently being written to will be closed and a new rowgroup will be created along with a
new dictionary. This has two significant implications:

• Dictionaries Consume Space: The more dictionaries that are needed

to support a columnstore index, and the larger they are, the more
storage space and memory are required to read and write to them.

• Undersized Rowgroups Create Fragmentation: A rowgroup may

contain up to 220 (1,048,576) rows. When rowgroups are undersized
by dictionary pressure (well below 220 rows), then more rowgroups,
segments, and dictionaries need to be maintained by SQL Server.
This in turn wastes resources and slows down analytics that need
these tables.

Fortunately, dictionary pressure is easy to diagnose. Once identified, a solution can

be crafted to resolve it. Consider the rowgroup metadata shown in Figure 15-5.

Figure 15-5. Example of dictionary pressure

257
Chapter 15 Columnstore Index Performance

In this example, row_group_id 25 shows a trim reason of DICTIONARY_SIZE and a

row count of 419,432 rows. In this rowgroup, SQL Server was loading rows into it when
the dictionary reached a size where no more entries could be added without bringing
it over the 16MB cap. If this were the only rowgroup like this, then dictionary pressure
may not be seen as a serious issue. If many rowgroups were cut off at row counts this low,
then dictionary pressure would without a doubt be inhibiting optimal performance and
would be worth investigating and resolving.
There are two primary causes of dictionary pressure:
• Columns that are very wide

• Columns with many distinct values

The most common scenarios involving dictionary pressure are derived from
columns that are both wide and have many distinct values. There is no one-size-fits-all
solution to the challenge of dictionary pressure. The three most common solutions can
be considered as the simplest ways of resolving this challenge:

• Normalize wide columns into a lookup (dimension) table.

• Sort on a different column.

• Partition the table.

Normalizing Wide Columns

The simplest solution requires schema changes and therefore may be disruptive, and
that is to normalize a wide column that is identified as the source of dictionary pressure.
Normalizing a column replaces a wide column (often string data) with a smaller numeric
column. The lookup table would likely be stored using a clustered rowstore index and a
supporting nonclustered rowstore index to handle queries against the text data.
When considering a columnstore index for a large analytic table, pay close attention
to any wide columns, and as part of testing, run through the exercise of creating a
columnstore index on the data. Once created, pay attention to rowgroup row counts and
if dictionary pressure is occurring or not. If rowgroups are undersized by 10% or more,
then consider normalizing the wide column if it is the culprit.
Lookup tables can contain other metadata that supports efficient searches against it,
such as hashes, search criteria, or details that can be indexed and used to quickly reduce
row counts before the text column needs to be interrogated.

258
Chapter 15 Columnstore Index Performance

Add or Change the Columnstore Sorting Column

Sometimes the way in which data is sorted within a columnstore index lends itself to
poor dictionary encoding. If a columnstore index is unsorted or is sorted on a column
that tends to group data together alongside nonsimilar values, then wider columns
might become the source of dictionary pressure.
The column that a columnstore index is ordered on should always be the one
that is most frequently used to filter queries. If there are a variety of common search
criteria, then there may be other sorting options that provide better compression and
performance.
This solution to dictionary pressure requires before-and-after testing of common
analytic queries to ensure that one problem (dictionary pressure) is not being traded for
another (poor analytic query performance).

Partitioning
One key element of partitioned columnstore indexes is that each partition contains its
own distinct columnstore index. For example, a table that is broken into ten partitions
will contain a separate columnstore structure in each partition. That includes separate
dictionaries, as well!
When columnstore index row counts become large (hundreds of millions or billions
of rows), partitioning can provide distinct performance and maintenance benefits in
SQL Server. In addition to partition elimination, an added performance improvement
can be seen in the separation of dictionaries across partitions. This solution is especially
effective when the values for wide columns change slowly over time. In these scenarios,
the data from one day to the next will have many repeated values, but across months or
years will not.
When partitioning a table, be sensitive to the size of the partitions that are created.
Because each represents a distinct, separate columnstore index, it is important that each
be large enough to benefit from the key properties of columnstore indexes. Therefore,
ensure that each partition has at least tens of millions of rows. Similarly, partitions
that are exceptionally large (billions of rows) may suffer the same problems as an
unpartitioned table with regard to dictionary pressure.
Testing is important in confirming the benefits of partitioning in a columnstore
index. Generally speaking, if a table contains hundreds of millions of rows (or more),
being able to subdivide it into smaller chunks using the columnstore ordering column

259
Chapter 15 Columnstore Index Performance

as the partitioning column will allow performance benefits to be realized. If dictionary

pressure exists, then this can become an effective way to manage it by splitting up
dictionaries across partitions.
Chapter 11 dives into partitioning in detail, including demonstrations of how to
implement it and how it can provide a host of benefits in addition to resolving dictionary
pressure challenges.
In general, dictionaries are very efficient at storing their lookup data and
columnstore indexes benefit greatly by using them. Dictionary pressure should be
seen as an out-of-the-ordinary situation and only investigated if it is causing blatant
performance problems.

Columnstore Indexes on Temporary Tables

For scenarios when analytic data needs to be crunched on the fly using temporary
data, there can be value in adding columnstore indexes to temporary objects. This is
a unique use case that will rarely be needed, but can provide exceptional value to ETL
processes or any analytics in which an intermediary or temporary data set requires many
expensive queries to be executed against it.
Adding a columnstore index to a temporary table works exactly as it does against a
permanent table. Listing 15-5 shows an example of how to accomplish this task.

Listing 15-5. Query That Creates a Temporary Table, Populates It, and Adds a
Columnstore Index
CREATE TABLE #Sales_Temp_Data
(      [Sale Key] BIGINT NOT NULL,
       [Customer Key] INT NOT NULL,
       [Invoice Date Key] DATE NOT NULL,
       Quantity INT NOT NULL,
       [Total Excluding Tax] DECIMAL(18,2) NOT NULL);

INSERT INTO #Sales_Temp_Data

([Sale Key], [Customer Key], [Invoice Date Key], Quantity, [Total
Excluding Tax])
SELECT

260
Chapter 15 Columnstore Index Performance

[Sale Key], [Customer Key], [Invoice Date Key], Quantity, [Total
Excluding Tax]
FROM Fact.Sale_CCI_PARTITIONED;

CREATE CLUSTERED COLUMNSTORE INDEX CCI_Sales_Temp_Data ON #Sales_Temp_Data;

Once added, the newly indexed temporary table may be queried as effectively as a
columnstore indexed permanent table, as seen in Listing 15-6.

Listing 15-6. Example Queries Against a Temporary Table with a

Columnstore Index
SELECT
COUNT(*)
FROM #Sales_Temp_Data;

SELECT
SUM(Quantity) * SUM([Total Excluding Tax])
FROM #Sales_Temp_Data
WHERE [Invoice Date Key] >= '1/1/2015'
AND [Invoice Date Key] < '1/1/2016';

These queries execute relatively quickly, returning the requested results. Figure 15-6
shows the STATISTICS IO output from that pair of queries.

Figure 15-6. IO for queries against a temporary table with a columnstore index

Note that the filtered query read almost half of the segments in the table, despite
only reading about a quarter of its rows. When this temporary table was populated, no
data order was enforced. As a result, data was inserted in whatever order SQL Server
happened to read it from the source table, which was not optimal. Depending on how

261
Chapter 15 Columnstore Index Performance

extensive the temporary table is to be used for analytics, taking the extra steps to order it
prior to reading it may or may not be worth the time and resources needed to do so. That
value will lie in whether or not the analytic processes are completing quickly enough and
if they are using too much computing resources along the way.
Nonclustered columnstore index may also be created on temporary tables, if there is
a need to slice data using both transactional and analytic methods. One benefit of doing
so is that the clustered rowstore index can enforce data order on the table, even if further
writes are made to it. Another benefit is that the column list for the columnstore index
can be customized. This allows a small subset of columns to be subject to analytics when
the remainder may be needed for other operations. Listing 15-7 shows the syntax for
creating nonclustered columnstore indexes on a temporary table.

Listing 15-7. Illustrating the Use of Nonclustered Columnstore Indexes on

Temporary Tables
CREATE TABLE #Sales_Temp_Data
(      [Sale Key] BIGINT NOT NULL,
       [Customer Key] INT NOT NULL,
       [Invoice Date Key] DATE NOT NULL,
       Quantity INT NOT NULL,
       [Total Excluding Tax] DECIMAL(18,2) NOT NULL);

INSERT INTO #Sales_Temp_Data

([Sale Key], [Customer Key], [Invoice Date Key], Quantity, [Total
Excluding Tax])
SELECT
[Sale Key], [Customer Key], [Invoice Date Key], Quantity, [Total
Excluding Tax]
FROM Fact.Sale_CCI_PARTITIONED;

CREATE CLUSTERED INDEX CI_Sales_Temp_Data ON #Sales_Temp_Data ([Sale Key]);

CREATE NONCLUSTERED COLUMNSTORE INDEX NCCI_Sales_Temp_Data ON

#Sales_Temp_Data ([Invoice Date Key], Quantity, [Total Excluding Tax]);

The syntax for nonclustered columnstore indexes is identical for temporary tables
and permanent tables, and once created, they can be used in the same fashion.

262
Chapter 15 Columnstore Index Performance

Note that columnstore indexes are not allowed on table variables. The T-SQL in
Listing 15-8 shows an attempt to do so.

Listing 15-8. A Script That Attempts to Add a Columnstore Index to a Table

Variable
DECLARE @Sales_Temp_Data TABLE
(      [Sale Key] BIGINT NOT NULL,
       [Customer Key] INT NOT NULL,
       [Invoice Date Key] DATE NOT NULL,
       Quantity INT NOT NULL,
       [Total Excluding Tax] DECIMAL(18,2) NOT NULL);

INSERT INTO @Sales_Temp_Data

CREATE CLUSTERED COLUMNSTORE INDEX CCI_Sales_Temp_Data ON @Sales_Temp_Data;

This query won’t even compile and will immediately generate the error shown in
Figure 15-7 when parsed.

Figure 15-7. Parsing error when trying to create a columnstore index on a table
variable

SQL Server provides the courtesy of raising an error during parsing, before the table
variable is created and populated.
Using columnstore indexes on temporary tables will not be a frequently applied use
case, but specific processes that rely heavily on the crunching of temporary data may be
able to use them effectively. Performance testing can be the ultimate test of whether this
is a helpful solution or one that does not provide enough value to be of worth.
263
Chapter 15 Columnstore Index Performance

Memory-Optimized Columnstore Indexes

A common use case for columnstore indexes is real-time operational analytics.
Nonclustered columnstore indexes provide the ability to execute analytic queries against
OLTP/rowstore tables efficiently. By being able to choose the columns to index, the
columnstore index can be used to target problematic OLAP queries in scenarios where
analytics cannot easily be off-loaded to other systems.
Memory-optimized tables are the ultimate in OLTP solutions. They provide
exceptional speed for heavily transactional workloads while eliminating the latching and
locking that typically accompanies contentious workloads. They can be targeted with
columnstore indexes, too, allowing analytics to be run against highly transactional tables
in memory. This is a niche use-case that can provide exceptional speeds for a select few
scenarios that will be discussed in this section.

Demonstrating Memory-Optimized Columnstore Indexes

Before diving into columnstore index queries, it is important to configure a database
with the necessary filegroup, file, and snapshot isolation setting needed to allow
memory-optimized tables to be created. Listing 15-9 provides the necessary
configuration changes to allow memory-optimized tables in SQL Server.

Listing 15-9. Enabling Memory-Optimized Tables in SQL Server

ALTER DATABASE WideWorldImporters ADD FILEGROUP WideWorldImporters_moltp

CONTAINS MEMORY_OPTIMIZED_DATA;
ALTER DATABASE WideWorldImporters ADD FILE (name='WideWorldImporters_moltp',
filename='C:\SQLData\WideWorldImporters_moltp') TO FILEGROUP
WideWorldImporters_moltp;
ALTER DATABASE WideWorldImporters SET
MEMORY_OPTIMIZED_ELEVATE_TO_SNAPSHOT = ON;

Only one memory-optimized filegroup is allowed per database. If a database

already has one, then trying to add another will result in an error like the one shown in
Figure 15-8.

264
Chapter 15 Columnstore Index Performance

Figure 15-8. Error encountered when adding a memory-optimized filegroup to a

database that already has one

With these basic configuration steps complete, we can experiment with a memory-
optimized table that contains a columnstore index. The script in Listing 15-10 shows the
creation of an example table.

Listing 15-10. Initial (Imperfect) Creation of a Memory-Optimized Table with a

Columnstore Index
CREATE TABLE Sales.Orders_MOLTP
(      OrderID INT NOT NULL CONSTRAINT PK_Orders_MOLTP PRIMARY KEY
NONCLUSTERED HASH WITH (BUCKET_COUNT = 150000),
       CustomerID INT NOT NULL,
       SalespersonPersonID INT NOT NULL,
       PickedByPersonID INT NULL,
       ContactPersonID INT NOT NULL,
       BackorderOrderID INT NULL,
       OrderDate DATE NOT NULL,
       ExpectedDeliveryDate DATE NOT NULL,
       CustomerPurchaseOrderNumber NVARCHAR(20) NULL,
       IsUndersupplyBackordered BIT NOT NULL,
       Comments NVARCHAR(MAX) NULL,
       DeliveryInstructions NVARCHAR(MAX) NULL,
       InternalComments NVARCHAR(MAX) NULL,
       PickingCompletedWhen DATETIME2(7) NULL,
       LastEditedBy INT NOT NULL,
       LastEditedWhen DATETIME2(7) NOT NULL,
       INDEX CCI_Orders_MOLTP CLUSTERED COLUMNSTORE)
WITH (MEMORY_OPTIMIZED = ON, DURABILITY = SCHEMA_AND_DATA);

265
Chapter 15 Columnstore Index Performance

This definition appears valid, but when executed, an error is returned, as shown in
Figure 15-9.

Figure 15-9. Error due to LOB columns included in memory-optimized

columnstore index definition

Columns of MAX length are not allowed in a memory-optimized columnstore index.

To allow its creation, these columns need to be normalized into another table or shrunk
to a smaller size. Listing 15-11 shows a revised script where the MAX length columns
have been reduced in size to accommodate this limitation.

Listing 15-11. Second (Imperfect) Creation of a Memory-Optimized Table with a

Columnstore Index
CREATE TABLE Sales.Orders_MOLTP
(      OrderID INT NOT NULL CONSTRAINT PK_Orders_MOLTP PRIMARY KEY
NONCLUSTERED HASH WITH (BUCKET_COUNT = 150000),
       CustomerID INT NOT NULL,
       SalespersonPersonID INT NOT NULL,
       PickedByPersonID INT NULL,
       ContactPersonID INT NOT NULL,
       BackorderOrderID INT NULL,
       OrderDate DATE NOT NULL,
       ExpectedDeliveryDate DATE NOT NULL,
       CustomerPurchaseOrderNumber NVARCHAR(20) NULL,
       IsUndersupplyBackordered BIT NOT NULL,
       Comments NVARCHAR(2000) NULL,
       DeliveryInstructions NVARCHAR(2000) NULL,
       InternalComments NVARCHAR(2000) NULL,
       PickingCompletedWhen DATETIME2(7) NULL,
       LastEditedBy INT NOT NULL,
       LastEditedWhen DATETIME2(7) NOT NULL,
       INDEX CCI_Orders_MOLTP CLUSTERED COLUMNSTORE)
WITH (MEMORY_OPTIMIZED = ON, DURABILITY = SCHEMA_AND_DATA);
266
Chapter 15 Columnstore Index Performance

With the NVARCHAR(MAX) columns replaced with NVARCHAR(2000), the script is

executed, resulting in a new error, as seen in Figure 15-10.

Figure 15-10. Error due to a memory-optimized columnstore index containing

more than 8060 bytes per row

While this error message appears to provide a glimmer of hope that this table definition
would work if it is executed on a version of SQL Server later than SQL Server 2014, the result
is still failure. After some review, a table definition is crafted that accommodates these
limitations and allows a table to (finally) be created, as seen in Listing 15-12.

Listing 15-12. Working Table Creation for a Memory-Optimized

Columnstore Index
CREATE TABLE Sales.Orders_MOLTP
(      OrderID INT NOT NULL CONSTRAINT PK_Orders_MOLTP PRIMARY KEY
NONCLUSTERED HASH WITH (BUCKET_COUNT = 150000),
       CustomerID INT NOT NULL,
       SalespersonPersonID INT NOT NULL,
       PickedByPersonID INT NULL,
       ContactPersonID INT NOT NULL,
       BackorderOrderID INT NULL,
       OrderDate DATE NOT NULL,
       ExpectedDeliveryDate DATE NOT NULL,
       CustomerPurchaseOrderNumber NVARCHAR(20) NULL,
       IsUndersupplyBackordered BIT NOT NULL,
       Comments NVARCHAR(500) NULL,
       DeliveryInstructions NVARCHAR(250) NULL,
       InternalComments NVARCHAR(500) NULL,
       PickingCompletedWhen DATETIME2(7) NULL,
       LastEditedBy INT NOT NULL,
       LastEditedWhen DATETIME2(7) NOT NULL,
       INDEX CCI_Orders_MOLTP CLUSTERED COLUMNSTORE)
WITH (MEMORY_OPTIMIZED = ON, DURABILITY = SCHEMA_AND_DATA);

267
Chapter 15 Columnstore Index Performance

Success! This table contains a handful of features that are not present on disk-
based tables:

• Nonclustered hash index is used as the primary key.

• MEMORY_OPTIMIZED = ON.

• DURABILITY = SCHEMA_AND_DATA.

The durability setting determines if this table’s data can be recovered when the
server is restarted or if only the schema is persisted. SCHEMA_ONLY is significantly
faster as there is no need to persist data to disk storage. In addition, startup is faster
as no data needs to be loaded into the table for it to be available. SCHEMA_ONLY is
typically used for tables that contain temporary data such as session, ETL, or transient
information that is not needed again once processed. SCHEMA_ONLY does not
support columnstore indexes, though, and therefore is out of scope for any further
discussion here.
Note that the columnstore index is labeled as a clustered columnstore index, but it
is not truly a clustered index. The primary storage mechanism for a memory-optimized
table is always the set of rows in memory. The columnstore index is an additional
structure on top of the memory-optimized object that is also persisted to disk storage.
These caveats incur quite a bit of overhead by SQL Server to maintain a columnstore
index alongside a memory-optimized table.
Also worthy of highlighting is the fact that this memory-optimized clustered
columnstore index contains numerous wide columns that are not ideal for dictionary
encoding. CustomerPurchaseOrderNumber, Comments, DeliveryInstructions, and
InternalComments are string columns that are unlikely to be repetitive. Therefore,
they are not likely to compress well and may cause dictionary pressure, resulting
in undersized rowgroups and further reduction in columnstore efficiency. This is
not a deal-breaker, but is essential to understand when considering implementing
a memory-optimized columnstore index. Tables that are built primarily for OLTP
workloads will often contain text data that is essential for transactional processing, but
that may be suboptimal for analytics. One possibility that could resolve a situation like
this would be to split the table into two, with the string columns and supporting data
in one table and the numbers and metrics in the other. This would allow one to be
given a memory-optimized columnstore index and the other to remain with its
memory-optimized rowstore structure.

268
Chapter 15 Columnstore Index Performance

Nonclustered columnstore indexes are not supported on memory-optimized

tables, and therefore the option to choose which columns to include in the index is not
available.
With a memory-optimized columnstore index created, the script in Listing 15-13 can
be executed to populate it with the same data as in the disk-based table Sales.Orders.

Listing 15-13. Script to Populate a Memory-Optimized Columnstore Index

with Data
INSERT INTO Sales.Orders_MOLTP
(      OrderID, CustomerID, SalespersonPersonID, PickedByPersonID,
ContactPersonID, BackorderOrderID, OrderDate, ExpectedDeliveryDate,
    CustomerPurchaseOrderNumber, IsUndersupplyBackordered, Comments,
DeliveryInstructions, InternalComments, PickingCompletedWhen,
    LastEditedBy, LastEditedWhen)
SELECT
       OrderID, CustomerID, SalespersonPersonID, PickedByPersonID,
ContactPersonID, BackorderOrderID, OrderDate, ExpectedDeliveryDate,
    CustomerPurchaseOrderNumber, IsUndersupplyBackordered, Comments,
DeliveryInstructions, InternalComments, PickingCompletedWhen,
    LastEditedBy, LastEditedWhen
FROM Sales.Orders;

With table creation and population complete, the sizes of the original Sales.Orders
table can be compared to its memory-optimized counterpart. Note the index contents of
each table:
Sales.Orders:

• A one-column primary key index

• Four single-column indexes

• One four-column index

Sales.Orders_MOLTP:
• A one-column primary key index

• A columnstore index (on all columns)

269
Chapter 15 Columnstore Index Performance

Figure 15-11. Size comparison of a disk-based table vs. a memory-optimized table

with a clustered columnstore index

Note in Figure 15-11 the significant size difference between the disk-based table
(25MB) and the memory-optimized table (117MB). That is a hefty space penalty and
underscores the fact that mapping a memory-optimized structure to a columnstore
structure is even more complex an operation than mapping a nonclustered rowstore
index to a clustered columnstore index. Before continuing, one additional
memory-optimized table will be created, as shown in Listing 15-14.

Listing 15-14. Table Creation for a Memory-Optimized Table Without a

Columnstore Index
CREATE TABLE Sales.Orders_MOLTP_NO_CCI
(      OrderID INT NOT NULL CONSTRAINT PK_Orders_MOLTP_NO_CCI PRIMARY KEY
NONCLUSTERED HASH WITH (BUCKET_COUNT = 150000),
       CustomerID INT NOT NULL,
       SalespersonPersonID INT NOT NULL,
       PickedByPersonID INT NULL,
       ContactPersonID INT NOT NULL,
       BackorderOrderID INT NULL,
       OrderDate DATE NOT NULL INDEX IX_Orders_MOLTP_NO_CCI_OrderDate
NONCLUSTERED,
       ExpectedDeliveryDate DATE NOT NULL,
       CustomerPurchaseOrderNumber NVARCHAR(20) NULL,
       IsUndersupplyBackordered BIT NOT NULL,
       Comments NVARCHAR(500) NULL,
       DeliveryInstructions NVARCHAR(250) NULL,

270
Chapter 15 Columnstore Index Performance

InternalComments NVARCHAR(500) NULL,

       PickingCompletedWhen DATETIME2(7) NULL,
       LastEditedBy INT NOT NULL,
       LastEditedWhen DATETIME2(7) NOT NULL)
WITH (MEMORY_OPTIMIZED = ON, DURABILITY = SCHEMA_AND_DATA);
GO

This table is identical to the previously created memory-optimized table, except that
the columnstore index has been swapped out for a nonclustered index on OrderDate.
The new table’s total size is 20MB, which is about one-sixth of the size of the columnstore
table, representing a significant space savings.
Consider a test analytic query that is executed against all three test tables, as shown
in Listing 15-15.

Listing 15-15. Test Analytic Query Against Each Orders Table

SELECT
COUNT(*) AS OrderCount,
COUNT(DISTINCT(CustomerID)) AS DistinctCustomerCount
FROM Sales.Orders
WHERE OrderDate >= '1/1/2015'
AND OrderDate < '4/1/2015';

SELECT
COUNT(*) AS OrderCount,
COUNT(DISTINCT(CustomerID)) AS DistinctCustomerCount
FROM Sales.Orders_MOLTP_NO_CCI
WHERE OrderDate >= '1/1/2015'
AND OrderDate < '4/1/2015';

SELECT
COUNT(*) AS OrderCount,
COUNT(DISTINCT(CustomerID)) AS DistinctCustomerCount
FROM Sales.Orders_MOLTP
WHERE OrderDate >= '1/1/2015'
AND OrderDate < '4/1/2015';

271
Chapter 15 Columnstore Index Performance

Results are returned from each query. The execution plans can be seen in
Figure 15-12.

Figure 15-12. Execution plans for a test analytic query against multiple tables

The disk-based execution plan suggests a covering index to avoid a clustered

index scan that incurs a total of 1383 reads. The memory-optimized table without a
columnstore index uses an index seek to return results, whereas the memory-optimized
table with a columnstore index uses a columnstore index scan to return the same results.
The columnstore index scan is by far the least expensive and fastest option out of the
three of these example tables. That being said, the cost of maintaining a memory-optimized
columnstore index is not trivial. There are a variety of limitations imposed on this structure,
beyond what has already been introduced:

• ALTER TABLE cannot be used on a memory-optimized columnstore

index to add columns, create indexes, change data types, or many
of the other common DDL changes made after a table is initially
created.

• The columnstore index is replicated to disk as a matter of necessity,

to help speed up database recovery after a server restart.

• Just as with disk-based tables, mixing clustered columnstore and

nonclustered rowstore indexes removes SQL Server’s ability to
automatically apply Vertipaq optimization.

272
Chapter 15 Columnstore Index Performance

• Computed columns are not allowed on memory-optimized

columnstore indexes.

• Because memory-optimized columnstore indexes are only allowed

as clustered indexes, filters may not be applied to them.

• Index maintenance is extremely limited. Rebuilds may only be

used to adjust hash index bucket counts (if needed) and reorganize
operations are not allowed. Sys.sp_memory_optimized_cs_migration
can be used to process the contents of the delta store, if needed.

• Columnstore archive compression is not allowed on

memory-optimized columnstore indexes.

The list of caveats related to memory-optimized columnstore indexes is significant

and makes it clear that this is not a feature that should be used without careful
consideration. The ideal scenario for putting a columnstore index is as follows:

• OLTP table with high concurrency.

• OLTP table that is identified as an exceptional candidate for

migration to a memory-optimized table.

• Table is targeted with frequent analytic queries that cause heavy

scans and unacceptable resource consumption.

• OLTP and OLAP operations cannot be split into separate data

sources.

Even if a table meets these criteria, be sure to test thoroughly and be able to quantify
that the performance benefits of the memory-optimized columnstore index outweigh
the limitations and drawbacks. This is a unique architectural feature that will only be
applicable to a small fraction of high-volume, highly concurrent OLTP tables. More often
than not, other alternatives will be more feasible, such as

• Disk-based table with a clustered rowstore index and a nonclustered

columnstore index.

• Memory-optimized table with additional covering nonclustered

indexes.

• Separate OLTP/OLAP data processing/storage systems.

273
Chapter 15 Columnstore Index Performance

Optimization Strategies
Generally speaking, optimization is a process that should begin with solutions that
address the most common use cases. Tweaking and implementing additional layers of
complexity should only occur when one or more of the following are true:

• Other solutions have failed to provide sufficient performance gains.

• The situation can be classified as an exception to the rules or an

edge case.

• Testing and QA of new strategies proves worthwhile gains (in

performance, resource savings, or both).

There is no one-size-fits-all solution to any set of architectural challenges, and

therefore it is important to use caution when applying techniques introduced in this
chapter (or anywhere for that matter). Keep it simple and only when simple fails should
less simple solutions be considered.

274
Index
A Brute-force approach, 17
Bulk insert processes, 45, 132
AlwaysOn Availability Groups, 183, 184
Bulk loading
Analytic data
columnstore indexes, 112, 114
characteristics, 18
performance, columnstore
OLAP vs. OLTP data access patterns, 16
indexes, 114–118
size, 2–4
processes, 111, 112
structure, 4, 5
Analytic query pattern, 101
Analytic sales order table, 3 C
Analytic workloads, 112, 179 Clustered columnstore index, 179, 207
compressed segments, 46
data storage, 24
B Clustered rowstore index, 23
Backups, 30, 111, 158, 160 data storage, 21
Batch mode execution structure, 46
columnstore aggregation query, 101 Columnstore archive compression,
execution plan 68–70, 244, 273
clustered index scan, 105 Columnstore compression
columnstore index, 100 algorithms
fact and dimension tables, 102 dictionary encoding, 54–60
performance discrepancies, 108 row order (vertipaq)
row counts, 105 optimization, 62–65
row mode vs. batch mode run-length encoding, 65, 66
operation, 109 string data normalization, 60, 62
performance, 107–110 value encoding, 51–54
query, 100 data stored by columns, 49
Batch mode processing, 98, 105–107 data structure, 50
Batch mode vs. row mode processing, 110 life cycle, 71
Binary tree structure, 37, 45 repeated values, 50
Bit array compression, 66, 67 repetitive columns, 49

275
© Edward Pollack 2022
E. Pollack, Analytics Optimization with Columnstore Indexes in Microsoft SQL Server,
https://doi.org/10.1007/978-1-4842-8048-5
INDEX

Columnstore indexes Compressed rowgroup, 42, 241, 242

aggregating single column, 25 Compression delay, nonclustered
analytic table, 25 columnstore indexes,
architectural components 189–195, 198
delete bitmap, 41 Constraints, 60, 212
delta store, 40, 41 Cost of modifying data, 123, 124
rowgroups and segments, 36–39 Covering indexes, 180, 182, 183, 204
sample data set, 33, 35, 36
benefits, 24
compression, 69 D
data stored, 24 Database maintenance tasks, 160
internal partitions, 93 Data growth, 28
loading data, practices, 118 Data load, 30
maintenance by partition, 246 columnstore reorganize
memory usage operations, 120–122
access_count, 91 drop nonclustered indexes, 119
object_load_time, 91, 92 Data movement/migration, 158, 159
object_type_desc, 90 Data reads, 253, 254
Sale_CCI Table, 89 Data set, 33, 35, 36
partitioning, 259 Data space, 149, 151
physical storage on pages, 43–47 Data warehouse, 6, 7
reorder and rebuild, 245, 246 Delete bitmap, 41, 125, 153, 225
vs. rowstore clustered index, 24 Deleted rows, 42, 75, 228, 229
SQL Server Delete operations
compression, 29, 30 columnstore index, 128
data loads, 30 delete bitmap, 125
faster analytic reads, 30 delta store, 125
native analytic data, 26, 27 rowgroup metadata, 126, 127
scalability, 27–29 rowgroups, 125
temporary tables, 260–263 Delta rowgroups, 94, 201, 227
Columnstore metadata, 143, 145 Delta store, 40, 41, 50, 118, 120–122, 134
rowgroup, 73–75 Dictionaries consume space, 257
rowgroup operational statistics, 84–88 Dictionary compression, 55, 57, 60
rowgroup physical, 79–84 Dictionary encoding, 54–60, 257
segment, 75–79 Dictionary metadata, 56, 58, 59
Columnstore segments, 67, 76–79, 90, 210 Dictionary pressure, 237, 239,
COMPRESS_ALL_ROW_GROUPS option, 241, 257–259
239, 240, 242 Disk-based execution plan, 272

276
INDEX

E L
Execution modes, 98, 101 Large objects (LOB), 67
Linear approximations, 29

F
M
Filegroup, 161
Memory-optimized columnstore
Filtered index, 195–198
indexes, 63
Filtered nonclustered columnstore
caveats, 273
indexes, 195–199
durability setting, 268
Filtered nonclustered rowstore
enabling, 264
indexes, 214, 215
error message, 267
Fragmentation
execution plans, test analytic query, 272
cause, 226
MAX length columns, 266
delete bitmap, 225
nonclustered columnstore indexes, 269
quantifying deleted rows, 228, 229
performance benefits, 273
unordered data, 230, 231, 233,
SCHEMA_ONLY, 268
234, 236
table creation, 265, 267, 270
unordered inserts, 227
test analytic query, 271
UPDATE operation, 227
Memory-optimized indexes, 25
Memory sizing, 254, 256
G Metadata reads, 249, 250, 252, 253
Microsoft Xpress compression algorithm, 69
Global dictionary, 56–58

N
H No-maintenance columnstore index, 236
Hot data, 43, 196 Nonclustered columnstore indexes
Huffman compression architecture, 42
algorithms, 69 benefits, 185
code changes, 199
compression delay, 188–194
I, J, K creation, 186
Indexed views, 218–220 DELETE and UPDATE operations, 185
Index maintenance, 160, 170, 247 execution plan, 186
INSERT operation, 132, 174, 236, 240 filtered, 195–199
Integer data compressed, 52, 53 hot, warm, and cold transactional
Internal columnstore index data, 187
objects, 93–95 index maintenance, 247
277
INDEX

Nonclustered columnstore challenge, 180

indexes (cont.) workloads, 18, 184
index usage statistics, 204, 205 Online order tracking system, 188
rowgroups, segments, and ONLINE rebuilds, 242, 243
compression, 185 Online transactional processing (OLTP)
STATISTICS IO, 187 challenge against rowstore tables, 180
testing, 203, 204 data and processes, 183
transactional table, 186 vs. OLAP data access patterns, 16
vertipaq optimization, 199, query retrieving, 22
201–203 storing analytic data, 7, 8
workloads, 186 workloads, 18
Nonclustered index, 23, 83, 95, 119 Operator properties
Nonclustered rowstore indexes columnstore and rowstore index
analytic query, 182, 208 scans, 102
benefits, 211 rowstore clustered index scan, 104
columnstore segment metadata, 210 Optimal analytic workloads, 154
compression, 220, 222 Optimization strategies, 274
enforcing constraints, 212, 214 Ordered data, 228, 233, 234
filtered, 214, 215 Over-indexing, 183
guidance, 222, 223
index creation query, 211
indexed views, 218–220
P
query execution plan and IO, 211 Page header, 44
Statistics IO output, 208 Page structure, SQL Server, 44
stock item key column, 209 Partition functions, 160, 165, 176
vertipaq optimization Partitioning
add filters, 216 add files to filegroup, 162
periodic index maintenance, boundary configuration, 165
216, 218 columnstore index maintenance, 246
Nonclustered rowstore primary configuration, 164
key, 98, 214 database files, 163
Normalization, 14, 60, 62 database maintenance, 160
Normalizing wide columns, 258 data modification, staging data, 174
data movement/migration, 158, 159
data types, 167
O elimination, 159
OLAP filegroup, 161
architects, 17 guidelines

278
INDEX

column choice, 176 invoice date key column,

partition and rowgroup sizing, 175 144, 149
storage choice, 176 metadata, 143
hot/warm/cold data, 157, 158 metadata query, 149
INSERT operation, 174 narrow analytic query, 146
narrow analytic queries, 169 ordered and unordered columnstore
partition scheme, creation, 165 index, 151
query execution plans, 170 power, 146
query optimizer, 170 row counts, 152
rebuild operation, 171 SQL Server, 153
staging table, switching operation, 172 STATISTICS IO, 145, 153
STATISTICS IO, 169 Rowgroup metadata, 240, 241
storage speed and cost, 158 columnstore index, 74, 127
table partitioning, benefits, 177 Deleted_rows, 75
table with ordered data set, 168 Delta_store_hobt_id, 75
Partition scheme, 165–167, 173, 250 Partition_number, 74
Partition switching, 161, 172, 175 Size_in_bytes, 75
Performance State and State_Description, 75
data reads, 253, 254 Total_rows, 75
memory sizing, 254, 256 Rowgroup operational statistics
metadata reads, 249, 250, 252, 253 columnstore index, query, 84
Periodic index maintenance, 216, 218 delete_buffer_scan_count, 86, 88
PowerPivot, 24, 63 index_scan_count, 85
returned_row_count, 88
scan_count, 85
Q Rowgroup physical metadata
Query execution plan, 105, 170, 211 created_time, 84
Query optimizer, 76, 159, 170 has_vertipaq_optimization, 83
query, 79
state_desc, 80
R sys.column_store_row_groups, 80
REBUILD operations, 171, 242, 244, 247 transition_to_compressed_state_
REORGANIZE operation, 237–240, desc, 82
242, 247 trim_reason_desc, 81, 82
Row batch structure, 105, 106 Rowgroups, 36–39
Rowgroup elimination, 155, 253 Rowgroup sizing, 175
data space, retrieve, 149 Row mode execution, 97–99
DELETE operations, 153 Row offsets, 44

279
INDEX

Row order (vertipaq) optimization, 62–65 T

Rowstore indexes, 47
Table partitioning, 157, 159, 177
analytic query, 180
Text-based data types, 4
covering index, 183
Third-party analytics software, 6
execution plan, analytic query, 182
Third-party tools, 184
OLAP and OLTP queries, 180
Time-based units, 145
STATISTICS IO, 181, 182
TOMBSTONE rowgroup, 202, 239, 241
threshold, 180
Transactional data, 11–14
Run-length encoded data, 124
Transactional data storage,
limits, 21–23
S Transactional order query, 13
Transactional queries, 22
Segment elimination, 78, 154, 253
Transactional sales order table, 3
columnstore index, 138
Transactional tables, 181, 187
filter and aggregate, 140
Transactional workloads, 18, 19, 264
optimization technique, 140
query, 140
query for two columns, 139 U
reducing resource consumption, 142
Unordered inserts, 227, 233, 235
rowstore tables, 140
Unstructured data, 1, 6
Segment metadata
Update operations, 128–135, 233, 247
base and magnitude, 78
columnstore index, 77
encoding_type, 77 V
has_nulls, 78 Value-based encoding, 77, 78
Min_data_id and Max_data_id, 78 Value compression, 53, 57, 60
null_value, 78 Value encoding, 51–54
on_disk_size, 79 Vertipaq compression, 63, 66
query, 76 Vertipaq optimization, 63, 65, 66, 83, 119,
row count, 78 124, 199, 201–203
Segment reads, 253
Segments, 36–39
SELECT query, 130 W
Sorting column, 259 Workloads, 18, 83, 97, 105
Spatial indexes, 25
SQL Server Analysis Services, 1, 24, 63
SQL Server storage, 22 X, Y, Z
Statistical metrics, 4 XML indexes, 25

280