Fast Data:

Smart and
at Scale
Design Patterns and Recipes

Ryan Betts & John Hugg

Make Data Work
strataconf.com
Presented by OReilly and Cloudera,
Strata + Hadoop World is where
cutting-edge data science and new
business fundamentals intersect
and merge.

n Learn business applications of

data technologies

n Develop new skills through

trainings and in-depth tutorials

n Connect with an international

community of thousands who
work with data

Job # 15420
 Fast Data:
Smart and at Scale
Design Patterns and Recipes

Ryan Betts and John Hugg

Fast Data: Smart and at Scale
by Ryan Betts and John Hugg
Copyright 2015 VoltDB, Inc. All rights reserved.
Printed in the United States of America.
Published by OReilly Media, Inc., 1005 Gravenstein Highway North, Sebastopol, CA
95472.
OReilly books may be purchased for educational, business, or sales promotional use.
Online editions are also available for most titles (http://safaribooksonline.com). For
more information, contact our corporate/institutional sales department:
800-998-9938 or [email protected].

Editor: Tim McGovern Interior Designer: David Futato

Production Editor: Dan Fauxsmith Cover Designer: Randy Comer
Illustrator: Rebecca Demarest

September 2015: First Edition

Revision History for the First Edition

2015-09-01: First Release
2015-10-20: Second Release

The OReilly logo is a registered trademark of OReilly Media, Inc. Fast Data: Smart
and at Scale, the cover image, and related trade dress are trademarks of OReilly
Media, Inc.
While the publisher and the authors have used good faith efforts to ensure that the
information and instructions contained in this work are accurate, the publisher and
the authors disclaim all responsibility for errors or omissions, including without
limitation responsibility for damages resulting from the use of or reliance on this
work. Use of the information and instructions contained in this work is at your own
risk. If any code samples or other technology this work contains or describes is sub
ject to open source licenses or the intellectual property rights of others, it is your
responsibility to ensure that your use thereof complies with such licenses and/or
rights.

978-1-491-94038-9
[LSI]
 Table of Contents

Foreword. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

Fast Data Application Value. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

Fast Data and the Enterprise. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

1. What Is Fast Data?. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

Applications of Fast Data 2
Uses of Fast Data 4

2. Disambiguating ACID and CAP. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

What Is ACID? 7
What Is CAP? 9
How Is CAP Consistency Different from ACID Consistency? 10
What Does Eventual Consistency Mean in This Context? 10

3. Recipe: Integrate Streaming Aggregations and Transactions. . . . . . 13

Idea in Brief 13
Pattern: Reject Requests Past a Threshold 14
Pattern: Alerting on Variations from Predicted Trends 14
When to Avoid This Pattern 15
Related Concepts 16

4. Recipe: Design Data Pipelines. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

Idea in Brief 17
Pattern: Use Streaming Transformations to Avoid ETL 18

iii
 Pattern: Connect Big Data Analytics to Real-Time Stream
Processing 19
Pattern: Use Loose Coupling to Improve Reliability 20
When to Avoid Pipelines 21

5. Recipe: Pick Failure-Recovery Strategies. . . . . . . . . . . . . . . . . . . . . . . 23

Idea in Brief 23
Pattern: At-Most-Once Delivery 24
Pattern: At-Least-Once Delivery 25
Pattern: Exactly-Once Delivery 26

6. Recipe: Combine At-Least-Once Delivery with Idempotent Processing

to Achieve Exactly-Once Semantics. . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Idea in Brief 27
Pattern: Use Upserts Over Inserts 28
Pattern: Tag Data with Unique Identifiers 29
Pattern: Use Kafka Offsets as Unique Identifiers 30
Example: Call Center Processing 31
When to Avoid This Pattern 32
Related Concepts and Techniques 33

Glossary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

iv | Table of Contents
 Foreword

We are witnessing tremendous growth of the scale and rate at which

data is generated. In earlier days, data was primarily generated as a
result of a real-world human actionthe purchase of a product, a
click on a website, or the pressing of a button. As computers become
increasingly independent of humans, they have started to generate
data at the rate at which the CPU can process ita furious pace that
far exceeds human limitations. Computers now initiate trades of
stocks, bid in ad auctions, and send network messages completely
independent of human involvement.
This has led to a reinvigoration of the data-management commu
nity, where a flurry of innovative research papers and commercial
solutions have emerged to address the challenges born from the
rapid increase in data generation. Much of this work focuses on the
problem of collecting the data and analyzing it in a period of time
after it has been generated. However, an increasingly important
alternative to this line of work involves building systems that pro
cess and analyze data immediately after it is generated, feeding
decision-making software (and human decision makers) with
actionable information at low latency. These fast data systems usu
ally incorporate recent research in the areas of low-latency data
stream management systems and high-throughput main-memory
database systems.
As we become increasingly intolerant of latency from the systems
that people interact with, the importance and prominence of fast
data will only grow in the years ahead.

Daniel Abadi, Ph.D.

Associate Professor, Yale University

v
 Fast Data Application Value

Looking Beyond Streaming

Fast data application deployments are exploding, driven by the
Internet of Things (IoT), a surge in data from machine-to-machine
communications (M2M), mobile device proliferation, and the reve
nue potential of acting on fast streams of data to personalize offers,
interact with customers, and automate reactions and responses.
Fast data applications are characterized by the need to ingest vast
amounts of streaming data; application and business requirements
to perform analytics in real time; and the need to combine the out
put of real-time analytics results with transactions on live data. Fast
data applications are used to solve three broad sets of challenges:
streaming analytics, fast data pipeline applications, and request/
response applications that focus on interactions.
While theres recognition that fast data applications produce signifi
cant valuefundamentally different value from big data applica
tionsits not yet clear which technologies and approaches should
be used to best extract value from fast streams of data.
Legacy relational databases are overwhelmed by fast datas require
ments, and existing tooling makes building fast data applications
challenging. NoSQL solutions offer speed and scale but lack transac
tionality and query/analytics capability. Developers sometimes stitch
together a collection of open source projects to manage the data
stream; however, this approach has a steep learning curve, adds
complexity, forces duplication of effort with hybrid batch/streaming
approaches, and limits performance while increasing latency.

vii
So how do you combine real-time, streaming analytics with real-
time decisions in an architecture thats reliable, scalable, and simple?
You could do it yourself using a batch/streaming approach that
would require a lot of infrastructure and effort; or you could build
your app on a fast, distributed data processing platform with sup
port for per-event transactions, streaming aggregations combined
with per-event ACID processing, and SQL. This approach would
simplify app development and enhance performance and capability.
This report examines how to develop apps for fast data, using well-
recognized, predefined patterns. While our expertise is with
VoltDBs unified fast data platform, these patterns are general
enough to suit both the do-it-yourself, hybrid batch/streaming
approach as well as the simpler, in-memory approach.
Our goal is to create a collection of fast data app development rec
ipes. In that spirit, we welcome your contributions, which will be
tested and included in future editions of this report. To submit a
recipe, send a note to [email protected].

viii | Fast Data Application Value

 Fast Data and the Enterprise

The world is becoming more interactive. Delivering information,

offers, directions, and personalization to the right person, on the
right device, at the right time and placeall are examples of new fast
data applications. However, building applications that enable real-
time interactions poses a new and unfamiliar set of data-processing
challenges. This report discusses common patterns found in fast
data applications that combine streaming analytics with operational
workloads.
Understanding the structure, data flow, and data management
requirements implicit in these fast data applications provides a
foundation to evaluate solutions for new projects. Knowing some
common patterns (recipes) to overcome expected technical hurdles
makes developing new applications more predictableand results
in applications that are more reliable, simpler, and extensible.
New fast data application styles are being created by developers
working in the cloud, IoT, and M2M. These applications present
unfamiliar challenges. Many of these applications exceed the scale of
traditional tools and techniques, creating new challenges not solved
by traditional legacy databases that are too slow and dont scale out.
Additionally, modern applications scale across multiple machines,
connecting multiple systems into coordinated wholes, adding com
plexity for application developers.
As a result, developers are reaching for new tools, new design tech
niques, and often are tasked with building distributed systems that
require different thinking and different skills than those gained from
past experience.

ix
This report is structured into four main sections: an introduction to
fast data, with advice on identifying and structuring fast data archi
tectures; a chapter on ACID and CAP, describing why its important
to understand the concepts and limitations of both in a fast data
architecture; four chapters, each a recipe/design pattern for writing
certain types of streaming/fast data applications; and a glossary of
terms and concepts that will aid in understanding these patterns.
The recipe portion of the book is designed to be easily extensible as
new common fast data patterns emerge. We invite readers to submit
additional recipes at [email protected]

x | Fast Data and the Enterprise

 CHAPTER 1
What Is Fast Data?

Into a world dominated by discussions of big data, fast data has been
born with little fanfare. Yet fast data will be the agent of change in
the information-management industry, as we will show in this
report.
Fast data is data in motion, streaming into applications and comput
ing environments from hundreds of thousands to millions of end
pointsmobile devices, sensor networks, financial transactions,
stock tick feeds, logs, retail systems, telco call routing and authoriza
tion systems, and more. Real-time applications built on top of fast
data are changing the game for businesses that are data dependent:
telco, financial services, health/medical, energy, and others. Its also
changing the game for developers, who must build applications to
handle increasing streams of data.1
Were all familiar with big data. Its data at rest: collections of struc
tured and unstructured data, stored in Hadoop and other data
lakes, awaiting historical analysis. Fast data, by contrast, is stream

1 Where is all this data coming from? Weve all heard the statement that data is doubling
every two yearsthe so-called Moores Law of data. And according to the oft-cited
EMC Digital Universe Study (2014), which included research and analysis by IDC, this
statement is true. The study states that data will multiply 10-fold between 2013 and
2020from 4.4 trillion gigabytes to 44 trillion gigabytes. This data, much of it new, is
coming from an increasing number of new sources: people, social, mobile, devices, and
sensors. Its transforming the business landscape, creating a generational shift in how
data is used, and a corresponding market opportunity. Applications and services tap
ping this market opportunity require the ability to process data fast.

1
ing data: data in motion. Fast data demands to be dealt with as it
streams in to the enterprise in real time. Big data can be dealt with
some other timetypically after its been stored in a Hadoop data
warehouseand analyzed via batch processing.
A stack is emerging across verticals and industries to help develop
ers build applications to process fast streams of data. This fast data
stack has a unique purpose: to process real-time data and output
recommendations, analytics, and decisionstransactionsin milli
seconds (billing authorization and up-sell of service level, for exam
ple, in telecoms), although some fast data use cases can tolerate up
to minutes of latency (energy sensor networks, for example).

Applications of Fast Data

Fast data applications share a number of requirements that influence
architectural choices. Three of particular interest are:

Rapid ingestion of millions of data eventsstreams of live data

from multiple endpoints
Streaming analytics on incoming data
Per-event transactions made on live streams of data in real time
as events arrive.

2 | Chapter 1: What Is Fast Data?

Ingestion
Ingestion is the first stage in the processing of streaming data. The
job of ingestion is to interface with streaming data sources and to
accept and transform or normalize incoming data. Ingestion marks
the first point at which data can be transacted against, applying key
functions and processes to extract value from datavalue that
includes insight, intelligence, and action.
Developers have two choices for ingestion. The first is to use direct
ingestion, where a code module hooks directly into the data-
generating API, capturing the entire stream at the speed at which the
API and the network will run, e.g., at wire speed. In this case, the
analytic/decision engines have a direct ingestion adapter. With
some amount of coding, the analytic/decision engines can handle
streams of data from an API pipeline without the need to stage or
cache any data on disk.
If access to the data-generating API is not available, an alternative is
using a message queue, e.g., Kafka. In this case, an ingestion system
processes incoming data from the queue. Modern queuing systems
handle partitioning, replication, and ordering of data, and can man
age backpressure from slower downstream components.

Streaming Analytics
As data is created, it arrives in the enterprise in fast-moving streams.
Data in a stream may arrive in many data types and formats. Most
often, the data provides information about the process that gener
ated it; this information may be called messages or events. This
includes data from new sources, such as sensor data, as well as click
streams from web servers, machine data, and data from devices,
events, transactions, and customer interactions.
The increase in fast data presents the opportunity to perform analyt
ics on data as it streams in, rather than post-facto, after its been
pushed to a data warehouse for longer-term analysis. The ability to
analyze streams of data and make in-transaction decisions on this
fresh data is the most compelling vision for designers of data-driven
applications.

Applications of Fast Data | 3

Per-Event Transactions
As analytic platforms mature to produce real-time summary and
reporting on incoming data, the speed of analysis exceeds a human
operators ability to act. To derive value from real-time analytics, one
must be able to take action in real time. This means being able to
transact against event data as it arrives, using real-time analysis in
combination with business logic to make optimal decisionsto
detect fraud, alert on unusual events, tune operational tolerances,
balance work across expensive resources, suggest personalized
responses, or tune automated behavior to real-time customer
demand.
At a data-management level, all of these actions mean being able to
read and write multiple, related pieces of data together, recording
results and decisions. It means being able to transact against each
event as it arrives.
High-speed streams of incoming data can add up to massive
amounts of data, requiring systems that ensure high availability and
at-least-once delivery of events. It is a significant challenge for enter
prise developers to create apps not only to ingest and perform ana
lytics on these feeds of data, but also to capture value, via per-event
transactions, from them.

Uses of Fast Data

Front End for Hadoop
Building a fast front end for Hadoop is an important use of fast data
application development. A fast front end for Hadoop should per
form the following functions on fast data: filter, dedupe, aggregate,
enrich, and denormalize. Performing these operations on the front
end, before data is moved to Hadoop, is much easier to do in a fast
data front end than it is to do in batch mode, which is the approach
used by Spark Streaming and the Lambda Architecture. Using a fast
front end carries almost zero cost in time to do filter, deduce, aggre
gate, etc., at ingestion, as opposed to doing these operations in a sep
arate batch job or layer. A batch approach would need to clean the
data, which would require the data to be stored twice, also introduc
ing latency to the processing of data.

4 | Chapter 1: What Is Fast Data?

An alternative is to dump everything in HDFS and sort it all out
later. This is easy to do at ingestion time, but its a big job to sort out
later. Filtering at ingestion time also eliminates bad data, data that is
too old, and data that is missing values; developers can fill in the val
ues, or remove the data if it doesnt make sense.
Then theres aggregation and counting. Some developers maintain
its difficult to count data at scale, but with an ingestion engine as the
fast front end of Hadoop its possible to do a tremendous amount of
counting and aggregation. If youve got a raw stream of data, say
100,000 events per second, developers can filter that data by several
orders of magnitude, using counting and aggregations, to produce
less data. Counting and aggregations reduce large streams of data
and make it manageable to stream data into Hadoop.
Developers also can delay sending aggregates to HDFS to allow for
late-arriving events in windows. This is a common problem with
other streaming systemsdata streams in a few seconds too late to a
window that has already been sent to HDFS. A fast data front end
allows developers to update aggregates when they come in.

Enriching Streaming Data

Enrichment is another option for a fast data front end for Hadoop.
Streaming data often needs to be filtered, correlated, or enriched
before it can be frozen in the historical warehouse. Performing
this processing in a streaming fashion against the incoming data
feed offers several benefits:

1. Unnecessary latency created by batch ETL processes is elimina

ted and time-to-analytics is minimized.
2. Unnecessary disk IO is eliminated from downstream big data
systems (which are usually disk-based, not memory-based,
when ETL is real time and not batch oriented).
3. Application-appropriate data reduction at the ingest point elim
inates operational expense downstreamless hardware is nec
essary.

The input data feed in fast data applications is a stream of informa

tion. Maintaining stream semantics while processing the events in
the stream discretely creates a clean, composable processing model.
Accomplishing this requires the ability to act on each input eventa

Uses of Fast Data | 5

capability distinct from building and processing windows, as is done
in traditional CEP systems.
These per-event actions need three capabilities: fast look-ups to
enrich each event with metadata; contextual filtering and sessioniz
ing (re-assembly of discrete events into meaningful logical events is
very common); and a stream-oriented connection to downstream
pipeline systems (e.g., distributed queues like Kafka, OLAP storage,
or Hadoop/HDFS clusters). This requires a stateful system fast
enough to transact on a per-event basis against unlimited input
streams and able to connect the results of that transaction process
ing to downstream components.

Queryable Cache
Queries that make a decision on ingest are another example of using
fast data front-ends to deliver business value. For example, a click
event arrives in an ad-serving system, and we need to know which
ad was shown, and analyze the response to the ad. Was the click
fraudulent? Was it a robot? Which customer account do we debit
because the click came in and it turns out that it wasnt fraudulent?
Using queries that look for certain conditions, we might ask ques
tions such as: Is this router under attack based on what I know
from the last hour? Another example might deal with SLAs: Is my
SLA being met based on what I know from the last day or two? If so,
what is the contractual cost? In this case, we could populate a dash
board that says SLAs are not being met, and it has cost n in the last
week. Other deep analytical queries, such as How many purple hats
were sold on Tuesdays in 2015 when it rained? are really best
served by systems such as Hive or Impala. These types of queries are
ad-hoc and may involve scanning lots of data; theyre typically not
fast data queries.

6 | Chapter 1: What Is Fast Data?

 CHAPTER 2
Disambiguating ACID and CAP

Fast data is transformative. The most significant uses for fast data
apps have been discussed in prior chapters. Key to writing fast data
apps is an understanding of two concepts central to modern data
management: the ACID properties and the CAP theorem, addressed
in this chapter. Its unfortunate that in both acronyms the C stands
for Consistency, but actually means completely different things.
What follows is a primer on the two concepts and an explanation of
the differences between the two C"s.

What Is ACID?
The idea of transactions, their semantics and guarantees, evolved
with data management itself. As computers became more powerful,
they were tasked with managing more data. Eventually, multiple
users would share data on a machine. This led to problems where
data could be changed or overwritten out from under users in the
middle of a calculation. Something needed to be done; so the aca
demics were called in.
The rules were originally defined by Jim Gray in the 1970s, and the
acronym was popularized in the 1980s. ACID transactions solve
many problems when implemented to the letter, but have been
engaged in a push-pull with performance tradeoffs ever since. Still,
simply understanding these rules can educate those who seek to
bend them.

7
A transaction is a bundling of one or more operations on database
state into a single sequence. Databases that offer transactional
semantics offer a clear way to start, stop, and cancel (or roll back) a
set of operations (reads and writes) as a single logical meta-
operation.
But transactional semantics do not make a transaction. A true
transaction must adhere to the ACID properties. ACID transactions
offer guarantees that absolve the end user of much of the headache
of concurrent access to mutable database state.
From the seminal Google F1 Paper:
The system must provide ACID transactions, and must always
present applications with consistent and correct data. Designing
applications to cope with concurrency anomalies in their data is
very error-prone, time-consuming, and ultimately not worth the
performance gains.

What Does ACID Stand For?

Atomic: All components of a transaction are treated as a single
action. All are completed or none are; if one part of a transac
tion fails, the databases state is unchanged.
Consistent: Transactions must follow the defined rules and
restrictions of the database, e.g., constraints, cascades, and trig
gers. Thus, any data written to the database must be valid, and
any transaction that completes will change the state of the data
base. No transaction will create an invalid data state. Note this is
different from consistency as defined in the CAP theorem.
Isolated: Fundamental to achieving concurrency control, isola
tion ensures that the concurrent execution of transactions
results in a system state that would be obtained if transactions
were executed serially, i.e., one after the other; with isolation, an
incomplete transaction cannot affect another incomplete trans
action.
Durable: Once a transaction is committed, it will persist and
will not be undone to accommodate conflicts with other opera
tions. Many argue that this implies the transaction is on disk as
well; most formal definitions arent specific.

8 | Chapter 2: Disambiguating ACID and CAP

What Is CAP?
CAP is a tool to explain tradeoffs in distributed systems. It was pre
sented as a conjecture by Eric Brewer at the 2000 Symposium on
Principles of Distributed Computing, and formalized and proven by
Gilbert and Lynch in 2002.

What Does CAP Stand For?

Consistent: All replicas of the same data will be the same value
across a distributed system.
Available: All live nodes in a distributed system can process
operations and respond to queries.
Partition Tolerant: The system will continue to operate in the
face of arbitrary network partitions.

The most useful way to think about CAP:

In the face of network partitions, you cant have both perfect con
sistency and 100% availability. Plan accordingly.
To be clear, CAP isnt about what is possible, but rather, what isnt
possible. Thinking of CAP as a You-Pick-Two theorem is misgui
ded and dangerous. First, picking AP or CP doesnt mean youre
actually going to be perfectly consistent or perfectly available; many
systems are neither. It simply means the designers of a system have
at some point in their implementation favored consistency or availa
bility when it wasnt possible to have both.
Second, of the three pairs, CA isnt a meaningful choice. The
designer of distributed systems does not simply make a decision to
ignore partitions. The potential to have partitions is one of the defi
nitions of a distributed system. If you dont have partitions, then you
dont have a distributed system, and CAP is just not interesting. If
you do have partitions, ignoring them automatically forfeits C, A, or
both, depending on whether your system corrupts data or crashes
on an unexpected partition.

What Is CAP? | 9
How Is CAP Consistency Different from ACID
Consistency?
ACID consistency is all about database rules. If a schema declares
that a value must be unique, then a consistent system will enforce
uniqueness of that value across all operations. If a foreign key
implies deleting one row will delete related rows, then a consistent
system will ensure the state cant contain related rows once the base
row is deleted.
CAP consistency promises that every replica of the same logical
value, spread across nodes in a distributed system, has the same
exact value at all times. Note that this is a logical guarantee, rather
than a physical one. Due to the speed of light, it may take some non
zero time to replicate values across a cluster. The cluster can still
present a logical view of preventing clients from viewing different
values at different nodes.
The most interesting confluence of these concepts occurs when sys
tems offer more than a simple key-value store. When systems offer
some or all of the ACID properties across a cluster, CAP consistency
becomes more involved. If a system offers repeatable reads,
compare-and-set or full transactions, then to be CAP consistent, it
must offer those guarantees at any node. This is why systems that
focus on CAP availability over CAP consistency rarely promise
these features.

What Does Eventual Consistency Mean in

This Context?
Lets consider the simplest case, a two-server cluster. As long as there
are no failures, writes are propagated to both machines and every
thing hums along. Now imagine the network between nodes is cut.
Any write to a node now will not propagate to the other node. State
has diverged. Identical queries to the two nodes may give different
answers.
The traditional response is to write a complex rectification process
that, when the network is fixed, examines both servers and tries to
repair and resynchronize state.
Eventual Consistency is a bit overloaded, but aims to address this
problem with less work for the developer. The original Dynamo

10 | Chapter 2: Disambiguating ACID and CAP

paper formally defined EC as the method by which multiple replicas
of the same value may differ temporarily, but would eventually con
verge to a single value. This guarantee that divergent data would be
temporary can render a complex repair and resync process unneces
sary.
EC doesnt address the issue that state still diverges temporarily,
allowing answers to queries to differ based on where they are sent.
Furthermore, EC doesnt promise that data will converge to the new
est or the most correct value (however that is defined), merely that it
will converge.
Numerous techniques have been developed to make development
easier under these conditions, the most notable being Conflict-free
Replicated Data Types (CRDTs), but in the best cases, these systems
offer fewer guarantees about state than CAP-consistent systems can.
The benefit is that under certain partitioned conditions, they may
remain available for operations in some capacity.
Its also important to note that Dynamo-style EC is very different
from the log-based rectification used by the financial industry to
move money between accounts. Both systems are capable of diverg
ing for a period of time, but the banks system must do more than
eventually agree; banks have to eventually have the right answer.
The next chapters provide examples of how to conceptualize and
write fast data apps.

What Does Eventual Consistency Mean in This Context? | 11

 CHAPTER 3
Recipe: Integrate Streaming
Aggregations and Transactions

Idea in Brief
Increasing numbers of high-speed transactional applications are
being built: operational applications that transact against a stream of
incoming events for use cases like real-time authorization, billing,
usage, operational tuning, and intelligent alerting. Writing these
applications requires combining real-time analytics with transaction
processing.
Transactions in these applications require real-time analytics as
inputs. Recalculating analytics from base data for each event in a
high-velocity feed is impractical. To scale, maintain streaming
aggregations that can be read cheaply in the transaction path. Unlike
periodic batch operations, streaming aggregations maintain consis
tent, up-to-date, and accurate analytics needed in the transaction
path.
This pattern trades ad hoc analytics capability for high-speed access
to analytic outputs that are known to be needed by an application.
This trade-off is necessary when calculating an analytic result from
base data for each transaction is infeasible.
Lets consider a few example applications to illustrate the concept.

13
Pattern: Reject Requests Past a Threshold
Consider a high-request-volume API that must implement sophisti
cated usage metrics for groups of users and individual users on a
per-operation basis. Metrics are used for multiple purposes: they are
used to derive usage-based billing charges, and they are used to
enforce a contracted quality of service standard (expressed as a
number of requests per second, per user, and per group). In this
case, the operational platform implementing the policy check must
be able to maintain fast counters for API operations, for users and
for groups. These counters must be accurate (they are inputs to bill
ing and quality of service policy enforcement), and they must be
accessible in real time to evaluate and authorize (or deny) new
requests.
In this scenario, it is necessary to keep a real-time balance for each
user. Maintaining the balance accurately (granting new credits,
deducting used credits) requires an ACID OLTP system. That same
system requires the ability to maintain high-speed aggregations.
Combining real-time, high-velocity streaming aggregations with
transactions provides a scalable solution.

Pattern: Alerting on Variations from Predicted

Trends
Imagine an operational monitoring platform that needs to issue
alerts or alarms when a threshold exceeds the predicated trend line
to a statistically significant level. This system combines two capabili
ties: it must maintain real-time analytics (streaming aggregations,
counters, and summary state of the current utilization), and it must
be able to compare these to the predicated trend. If the trend is
exceeded, the system must generate an alert or alarm. Likely, the sys
tem will record this alarm to suppress an alarm storm (to throttle
the rate of alarm publishing for a singular event).
This is another system that requires the combination of analytical
and transactional capability. Without the combined capability, this
problem would need three separate systems working in unison: an
analytics system that is micro-batching real-time analytics; an appli
cation reading those analytics and reading the predicated trendline
to generate alerts and alarms; and a transactional system that is stor
ing generated alert and alarm data to implement the suppression

14 | Chapter 3: Recipe: Integrate Streaming Aggregations and Transactions

logic. Running three tightly coupled systems like this (the solution
requires all three systems to be running) lowers reliability and com
plicates operations.

Figure 3-1. Streaming aggregations with transactions

Combining streaming event processing with request-response style

applications allows operationalizing real-time analytics.

When to Avoid This Pattern

Traditional OLAP systems offer the benefit of fast analytic queries
without pre-aggregation. These systems can execute complex quer
ies that scan vast amounts of data in seconds to minutesinsuffi
cient for high-velocity event feeds but within the threshold for many

When to Avoid This Pattern | 15

batch reporting functions, data science, data exploration, and
human analyst workflows. However, these systems do not support
high-velocity transactional workloads. They are optimized for
reporting, not for OLTP-style applications.

Related Concepts
Pre-aggregation is a common technique for which many algorithms
and features have been developed. Materialized views, probabilistic
data structures (examples: HyperLogLog, Bloom filters), and win
dowing are common techniques for implementing efficient real-
time aggregation and summary state.

1. Materialized Views: a view defines an aggregation, partitioning,

filter or join, grouping of base data. Materialized views main
tain a physical copy of the resulting tuples. Materialized Views
allow declarative aggregations, eliminating user code and ena
bling succinct, correct, and easy aggregations.
2. Probabilistic data structures aggregate data within some proba
bilistically bounded margin of error. These algorithms typically
trade precision for space, enabling bounded estimation in a
much smaller storage footprint. Examples of probabilistic data
structures include Bloom filters and HyperLogLog algorithms.
3. Windows are used to express moving averages, or time-
windowed summaries of a continuous event timeline. These
techniques are often found in CEP style or micro-batching sys
tems. SQL analytic functions (OVER, PARTITION) bring this
functionality to SQL platforms.

16 | Chapter 3: Recipe: Integrate Streaming Aggregations and Transactions

 CHAPTER 4
Recipe: Design Data Pipelines

Idea in Brief
Processing big data effectively often requires multiple database
engines, each specialized to a purpose. Databases that are very good
at event-oriented real-time processing are likely not good at batch
analytics against large volumes. Some systems are good for high-
velocity problems. Others are good for large-volume problems.
However, in most cases, these systems need to interoperate to sup
port meaningful applications.
Minimally, data arriving at the high-velocity, ingest-oriented sys
tems needs to be processed and captured into the volume-oriented
systems. In more advanced cases, reports, analytics, and predictive
models generated from the volume-oriented systems need to be
communicated to the velocity-oriented system to support real-time
applications. Real-time analytics from the velocity side need to be
integrated into operational dashboards or downstream applications
that process real-time alerts, alarms, insights, and trends.
In practice, this means that many big data applications sit on top of
a platform of tools. Usually the components of the platform include
at least a large shared storage pool (like HDFS), a high-performance
BI analytics query tool (like a columnar SQL system), a batch pro
cessing system (MapReduce or perhaps Spark), and a streaming sys
tem. Data and processing outputs move between all of these
systems. Designing that dataflowdesigning a processing pipeline

17
that coordinates these different platform components, is key to
solving many big data challenges.

Pattern: Use Streaming Transformations to

Avoid ETL
New events being captured into a long-term repository often require
transformation, filtering, or processing before they are available for
reporting use cases. For example, many applications capture sessions
comprising several discrete events, enrich events with static dimen
sion data to avoid expensive repeated joins in the batch layer, or fil
ter redundant events from a dataset storing only unique values.
There are at least two approaches to running these transformations.

1. All of the data can be landed to a long-term repository and then

extracted, transformed, and re-loaded back in its final form.
This approach trades I/O, additional storage requirements, and
longer time to insight (reporting is delayed until the ETL pro
cess completes) for a slightly simpler architecture (likely data is
moving directly from a queue to HDFS). This approach is
sometimes referred to as schema-on-read. It narrows the choice
of backend systems to those systems that are relatively schema-
freesystems that may not be optimal, depending on your spe
cific reporting requirements.
2. The transformations can be executed in a streaming fashion
before the data reaches the long-term repository. This approach
adds a streaming component between the source queue and the
final repository, creating a continuous processing pipeline.
Moving the transformation to a real-time processing pipeline
has several advantages. The write I/O to the backend system is
at least halved (in the first model, first raw data is written and
then ETLd data is written. In this model, only ETLd data is writ
ten.) This leaves more I/O budget available for data science and
reporting activitythe primary purpose of the backend reposi
tory. Operational errors are noticeable in almost real time.
When using ETL, raw data is not inspected until the ETL pro
cess runs. This delays operational notifications of missing or
corrupt inputs. Finally, time-to-insight is reduced. For example,
when organizing session data, a session is available to partici
pate in batch reporting immediately upon being closed. When

18 | Chapter 4: Recipe: Design Data Pipelines

 ETLing, you must wait, on average, for half of the ETL period
before the data is available for backend processing.

Pattern: Connect Big Data Analytics to Real-

Time Stream Processing
Real-time applications processing incoming events often require
analytics from backend systems. For example, if writing an alerting
system that issues notifications when a moving five minute interval
exceeds historical patterns, the data describing the historical pattern
needs to be available. Likewise, applications managing real-time cus
tomer experience or personalization often use customer segmenta
tion reports generated by statistical analysis run on the batch
analytics system. A third common example is hosting OLAP outputs
in a fast, scalable query cache to support operators and applications
that need high-speed, highly concurrent access to data.
In many cases, the reports, analytics, or models from big data ana
lytics need to be made available to real-time applications. In this
case, information flows from the big data (batch) oriented system to
the high-velocity (streaming) system. This introduces a few impor
tant requirements. First, the fast data, velocity-oriented application
requires a data management system capable of holding the state gen
erated by the batch system; second, this state needs to be regularly
updated or replaced in full. There are a few common ways to man
age the refresh cyclethe best tradeoff will depend on your specific
application.
Some applications (for example, applications based on user segmen
tation) require per-record consistency but can tolerate eventual con
sistency across records. In these cases, updating state in the velocity-
oriented database on a per-record basis is sufficient. Updates will
need to communicate new records (in this example, new custom
ers), updates to existing records (customers that have been recatego
rized), and deletions (ex-customers). Records in the velocity system
should be timestamped for operational monitoring and alerts gener
ated if stale records persist beyond the expected refresh cycle.
Other applications require the analytics data to be strictly consistent;
if it is insufficient for each record to be internally consistent, the set
of records as a whole requires a consistency guarantee. These cases
are often seen in analytic query caches. Often these caches are quer

Pattern: Connect Big Data Analytics to Real-Time Stream Processing | 19

ied for additional levels of aggregation, aggregations that span mul
tiple records. Producing a correct result therefore requires that the
full data set be consistent. A reasonable approach to transferring
these report data from the batch analytics system to the real-time
system is to write the data to a shadow table. Once the shadow table
is completely written, it can be atomically renamed, or swapped,
with the main table that is addressed by the application. The appli
cation will either see only data from the previous version of the
report or only data from the new version of the report, but will
never see a mix of data from both reports in a single query.

Pattern: Use Loose Coupling to Improve

Reliability
When connecting multiple systems, it is imperative that all systems
have an independent fate. Any part of the pipeline should be able to
fail while leaving other systems available and functional. If the batch
backend is offline, the high-velocity front end should still be operat
ing, and vice versa. This requires thinking through several design
decisions:

1. Where does data that can not be pushed (or pulled) through the
pipeline rest? Which components are responsible for the dura
bility of stalled data?
2. How do systems recover? Which systems are systems of record
meaning they can be recovery sources for lost data or inter
rupted processing?
3. What is the failure and availability model of each component in
the pipeline?
4. When a component fails, which other components become
unavailable? How long can upstream components maintain
functionality (for example, how long can they log processed
work to disk)? These numbers inform your recovery time objec
tive (RTO).

In every pipeline, there is by definition a slowest componenta bot

tleneck. When designing, explicitly choose the component that will
be your bottleneck. Having many systems, each with identical per
formance, means a minor degradation to any system will create a
new overall bottleneck. This is operationally painful. It is often bet

20 | Chapter 4: Recipe: Design Data Pipelines

ter to choose your most reliable component as your bottleneck or
your most expensive resource as your bottleneck. Overall you will
achieve a more predictable level of reliability.

When to Avoid Pipelines

Tying multiple systems together is always complex. This complexity
is rarely linear with the number of systems. Typically, complexity
increases as a function of connections (and worst case you can have
N^2 connections between N systems). To the extent your problem
can be fully solved by a single stand-alone system, you should prefer
that approach. However, large-volume and high-velocity data man
agement problems typically require a combination of specialized
systems. When combining these systems, carefully, consciously
design dataflow and connections between them. Loosely couple sys
tems so each operates independently of the failure of its connected
partners. Use multiple systems to simplify when possiblefor exam
ple, by moving batch ETL processes to continuous processes.

When to Avoid Pipelines | 21

 CHAPTER 5
Recipe: Pick Failure-Recovery
Strategies

Idea in Brief
Most streaming applications move data through multiple processing
stages. In many cases, events are landed in a queue and then read by
downstream components. Those components might write new data
back to a queue as they process or they might directly stream data to
their downstream components. Building a reliable data pipeline
requires designing failure-recovery strategies.
With multiple processing stages connected, eventually one stage will
fail, become unreachable, or otherwise unavailable. When this
occurs, the other stages continue to receive data. When the failed
component comes back online, typically it must recover some previ
ous state and then begin processing new events. This recipe dis
cusses where to resume processing.

The idempotency recipe discusses a specific

technique to achieve exactly-once semantics.

Additionally, for processing pipelines that are horizontally scaled,

where each stage has multiple servers or process running in parallel,
a subset of servers within the cluster can fail. In this case, the failed
server needs to be recovered, or its work needs to be reassigned.

23
There are a few factors that complicate these problems and lead to
different trade-offs.
First, it is usually uncertain what the last processed event was. It is
typically not technologically feasible, for example, to two-phase
commit the event processing across all pipeline components. Typi
cally, unreliable communication between processing components
means the fate of events in-flight near the time of failure is
unknown.
Second, event streams are often partitioned (sharded) across a num
ber of processors. Processors and upstream sources can fail in arbi
trary combinations. Picturing a single, unified event flow is often an
insufficient abstraction. Your system should assume that a single
partition of events can be omitted from an otherwise available
stream due to failure.
This leads to three options for resuming processing distinguished by
how events near the failure time are handled. Approaches to solving
the problem follow.

Pattern: At-Most-Once Delivery

At-most-once delivery allows some events to be dropped. In these
cases, events not processed because of an interruption are simply
dropped. They do not become part of the input to the downstream
system. If the data itself is low value or loses value if it is not imme
diately processed, this may be acceptable.
Questions you should answer when considering at-most-once deliv
ery include:

Will historical analytics know that data was unavailable?

Will the event stream eventually be archived to an OLAP store
or data lake? If so, how will future reporting and data science
algorithms detect and manage the missing values?
Is it clear what data was lost?

Remember that lost data is unlikely to align exactly with session

boundaries, time windows, or even data sources. It is also likely that
only partial data was dropped during the outage periodmeaning
that some values are present.

24 | Chapter 5: Recipe: Pick Failure-Recovery Strategies

Additional questions to be answered include: Is there a maximum
period of outage? What is the largest gap that can be dropped?
If the pipeline is designed to ignore events during outages, you
should determine the mean time to recovery for each component of
the pipeline to understand what volume of data will be lost in a typi
cal failure. Understanding the maximum allowable data loss should
be an explicit consideration when planning an at-most-once deliv
ery pipeline.
Many data pipelines are shared infrastructure. The pipeline is a plat
form that supports many applications. You should consider whether
all current and all expected future applications can detect and toler
ate data loss due to at-most-once delivery.
Finally, it is incorrect to assume that at-most-once delivery is the
default if another strategy is not explicitly chosen. You should not
assume that data during an outage is always discarded by upstream
systems. Many systems, queues especially, checkpoint subscriber
read points and resume event transmission from the checkpoint
when recovering from a failure. (This is actually an example of at-
least-once delivery.) Designing at-most-once delivery requires
explicit choices and implementationit is not the free choice.

Pattern: At-Least-Once Delivery

At-least-once delivery replays recent events starting from a known-
processed (acknowledged) event. This approach presents some data
to the processing pipeline more than once. The typical implementa
tion backing at-least-once delivery checkpoints a safe-point (that is
known to have been processed). After a failure, processing resumes
from the checkpoint. It is likely that events were successfully pro
cessed after the checkpoint. These events will be replayed during
recovery. This replay means that downstream components see each
event at least once.
There are a number of considerations when using at-least- once
delivery.
At-least-once delivery can lead to out-of-order event delivery. In
reality, regardless of the failure model chosen, you should assume
that some events will arrive late, out of order, or not at all.

Pattern: At-Least-Once Delivery | 25

Data sources are not well coordinated and rarely are events from
sources delivered end-to-end over a single TCP/IP connection (or
some other order-guaranteeing protocol).
If processing operations are not idempotent, replaying events will
corrupt or change outputs. When designing at-least-once delivery,
identify and characterize processes as idempotent or not.
If processing operations are not deterministic, replaying events will
produce different outcomes. Common examples of non-
deterministic operations include querying the current wallclock
time or invoking a remote service (that may be unavailable).
At-least-once delivery requires a durability contract with upstream
components. In the case of failure, some upstream component must
have a durable record of the event from which to replay. You should
clearly identify durability responsibility through the pipeline and
manage and monitor durable components appropriately, testing
operational behavior when disks fail or fill.

Pattern: Exactly-Once Delivery

Exactly-once processing is the idealeach event is processed exactly
once. This avoids the difficult side effects and considerations raised
by at-most-once and at-least-once processing. See Chapter 6 for
strategies on achieving exactly-once semantics using idempotency in
combination with at-least-once delivery.
Understanding that input streams are typically partitioned across
multiple processors, that inputs can fail on a per-partition basis, and
that events can be recovered using different strategies are all funda
mental aspects of designing distributed recovery schemes.

26 | Chapter 5: Recipe: Pick Failure-Recovery Strategies

 CHAPTER 6
Recipe: Combine At-Least-Once
Delivery with Idempotent
Processing to Achieve Exactly-Once
Semantics

Idea in Brief
When dealing with streams of data in the face of possible failure,
processing each datum exactly once is extremely difficult. When the
processing system fails, it may not be easy to determine which data
was successfully processed and which data was not.
Traditional approaches to this problem are complex, require
strongly consistent processing systems, and require smart clients
that can determine through introspection what has or hasnt been
processed.
As strongly consistent systems have become more scarce, and
throughput needs have skyrocketed, this approach often has been
deemed unwieldy and impractical. Many have given up on precise
answers and chosen to work toward answers that are as correct as
possible under the circumstances. The Lambda Architecture propo
ses doing all calculations twice, in two different ways, to allow for
cross-checking. Conflict-free replicated data types (CRDTs) have
been proposed as a way to add data structures that can be reasoned
about when using eventually consistent data stores.

27
If these options are less than ideal, idempotency offers another path.
An idempotent operation is an operation that has the same effect no
matter how many times it is applied. The simplest example is setting
a value. If I set x = 5, then I set x = 5 again, the second action doesnt
have any effect. How does this relate to exactly-once processing? For
idempotent operations, there is no effective difference between at-
least-once processing and exactly-once processing, and at-least-once
processing is much easier to achieve.
Leveraging the idempotent setting of values in eventually consistent
systems is one of the core tools used to build robust applications on
these platforms. Nevertheless, setting individual values is a much
weaker tool than the ACID-transactional model that pushed data
management forward in the late 20th century. CRDTs offer more,
but come with rigid constraints and restrictions. Theyre still a dan
gerous thing to build around without a deep understanding of what
they offer and how they work.
With the advent of consistent systems that truly scale, a broader set
of idempotent processing can be supported, which can improve and
simplify past approaches dramatically. ACID transactions can be
built that read and write multiple values based on business logic,
while offering the same effects if repeatedly executed.

Pattern: Use Upserts Over Inserts

An upsert is shorthand for describing a conditional insert; if the row
exists, dont insert; if the row does not, insert it. Some systems sup
port specific syntax for this. In SQL, this can involve an ON CON
FLICT clause, a MERGE statement, or even a straightforward
UPSERT statement. Some NoSQL systems have ways to express
the same thing. For key-value stores, the default behavior of put is
an upsert. Check your systems documentation.
When dealing with rows that can be uniquely identified, either
through a unique key or a unique value, upsert is a trivially idempo
tent operation.
When the status of an upsert is unclear, often due to client, server
node, or network failure, it is safe to send repeatedly until its success
can be verified. Note that this type of retry often should make use of
exponential backoff.

28 | Chapter 6: Recipe: Combine At-Least-Once Delivery with Idempotent Processing to

Achieve Exactly-Once Semantics
Pattern: Tag Data with Unique Identifiers
Idempotent operations are more difficult when data isnt uniquely
identifiable. Imagine a digital ad-tech app that tracks clicks on a web
page. An event arrives as a three-tuple that says user X clicked on
spot Y at time T (with second resolution). With this design, the
upsert pattern cant be used because it would be possible to record
multiple clicks by the same user in the same spot in the same sec
ond, e.g., a double-click.

Subpattern: Fine-Grained Timestamps

One solution to the non-unique clicks problem is to increase the
timestamp resolution to a point at which clicks are unique. If the
timestamp stored milliseconds, it might be reasonable to assume
that a user couldnt click faster than once per millisecond. This
makes upsert usable and idempotency attainable.
Note that its critical to verify on the client side that generated events
are in fact unique. Trusting a computer time API to accurately
reflect real-world time is a common mistake, proven time and again
to be dangerous. For example, some hardware/software/APIs offer
millisecond values, but 100 ms resolution. NTP (network time pro
tocol) is able to move clocks backward in many default configura
tions. Virtualization software is notorious for messing with guest OS
clocks.
To do this well, check that the last event and the new event have dif
ferent times on the client side before sending the event to the server.

Subpattern: Unique IDs at the Event Source

If you can generate a unique id at the client, send that value with the
event tuple to ensure that it is unique. If events are generated in one
place, its possible that a simple incrementing counter can uniquely
identify events. The trick with a counter is to ensure you dont re-use
values after restarting some service.
One approach is to use a central server to dole out blocks of unique
ids. A database with strong consistency or an agreement system
such as ZooKeeper can be used to assign blocks of ten thousand, one
hundred thousand, or one million ids in chunks. If the event pro
ducer fails, then some ids are wasted, but 64 bits should have
enough ids to cover any loss.

Pattern: Tag Data with Unique Identifiers | 29

Another approach is to combine timestamps with ids for unique
ness. Are you using millisecond timestamps but want to ensure
uniqueness? Start a new counter for every millisecond. If two events
share a millisecond, give one counter value 0 and another counter
value 2. This ensures uniqueness.
Another approach is to combine timestamps and counters in a 64-
bit number. VoltDB generates unique ids dividing a 64-bit integer
into sections, using 41 bits to identify a millisecond timestamp, 10
bits as a per-millisecond counter, and 10 bits as an event source id.
This leaves one bit for the sign, avoiding issues mixing signed and
unsigned integers. Note that 41 bits of milliseconds is about 70
years. You can play with the bit sizes for each field as needed. Be
very careful to anticipate and handle the case where time moves
backward or stalls.
If youre looking for something conceptually simpler than getting
incrementing ids correct, try an off-the-shelf UUID library to gener
ate universally unique IDs. These work in different ways, but often
combine machine information, such as a MAC address, with ran
dom values and timestamp values, similar to what is described
above. The upside is that it is safe to assume UUIDs are unique
without much effort, but the downside is they often require 16 or
more bytes to store.

Pattern: Use Kafka Offsets as Unique

Identifiers
Unique identifiers are built-in when using Kafka. Combining the
topic ID with the offset in the log can uniquely identify the event.
This sounds like a slam dunk, but there are reasons to be careful.

1. Inserting items into Kafka has all of the same problems as any
other distributed system. Managing exactly-once insertion into
Kafka is not easy, and Kafka doesnt offer the right tools (at this
time) to manage idempotency when writing to the Kafka topic.
2. If the Kafka cluster is restarted or switched, topic offsets may no
longer be unique. It may be possible to use a third value, e.g., a
Kafka cluster ID, to make the event unique.

30 | Chapter 6: Recipe: Combine At-Least-Once Delivery with Idempotent Processing to

Achieve Exactly-Once Semantics
Example: Call Center Processing
Consider a customer support call center with two events:

1. A caller is put on hold (or starts a call).

2. A caller is connected to an agent.

The app must ingest these events and compute average hold time
globally.

Version 1: Events Are Ordered

In this version, events for a given customer always arrive in the
order in which they are generated. Event processing in this example
is idempotent, so an event may arrive multiple times, but it can
always be assumed that events arrive in the order they happened.
The schema for state contains a single tuple containing the total
hold time and the total number of hold occurrences. It also contains
a set of ongoing holds.
When a caller is put on hold (or a call is started), upsert a record
into the set of ongoing holds. Use one of the methods described
above to assign a unique id. Using an upsert instead of an insert
makes this operation idempotent.
When a caller is connected to an agent, look up the corresponding
ongoing hold in the state. If the ongoing hold is found, remove it,
calculate the duration based on the two correlated events, and
update the global hold time and global hold counts accordingly. If
this message is seen repeatedly, the ongoing hold record will not be
found the second time it will be processed and can be ignored at that
point.
This works quite well, is simple to understand, and is space efficient.
But is the key assumption valid? Can we guarantee order? The
answer is that guaranteeing order is certainly possible, but its hid
den work. Often its easier to break the assumption on the process
ing end, where you may have an ACID-consistent processor that
makes dealing with complexity easier.

Example: Call Center Processing | 31

Version 2: Events Are Not Ordered
In this version, events may arrive in any order. The problem with
unordered events is that you cant delete from the outstanding holds
table when you get a match. What you can do, in a strongly consis
tent system, is keep one row per hold and mark it as matched when
its duration is added to the global duration sum. The row must be
kept around to catch any repeat messages.
How long do we need to keep these events? We must hold them
until were sure that another event for a particular hold could not
arrive. This may be minutes, hours, or days, depending on your sit
uation.
This approach is also simple, but requires additional state. The cost
of maintaining additional state should be weighed against the value
of perfectly correct data. Its also possible to drop event records early
and allow data to be slightly wrong, but only when events are
delayed abnormally. This may be a decent compromise between
space and correctness in some scenarios.

When to Avoid This Pattern

Idempotency can add storage overhead to store extra IDs for
uniqueness. It can add a fair bit of complexity, depending on many
factors, such as whether your event processor has certain features or
whether your app requires complex operations to be idempotent.
Making the effort to build an idempotent application should be
weighed against the cost of having imperfect data once in a while. Its
important to keep in mind that some data has less value than other
data, and spending developer time ensuring its perfectly processed
may be a poor allocation of resources.
Another reason to avoid idempotent operations is that the event
processor or data store makes it very hard to achieve, based on the
functionality you are trying to deliver, and switching tools is not a
good option.

32 | Chapter 6: Recipe: Combine At-Least-Once Delivery with Idempotent Processing to

Achieve Exactly-Once Semantics
Related Concepts and Techniques
Delivery Guarantees: discussed in detail in Chapter 5
Exponential Backoff: defined in the Glossary.
CRDTs: defined in the Glossary and referenced in Idea in
Brief on page 27.
ACID: defined and discussed in detail in Chapter 2.

Related Concepts and Techniques | 33

 Glossary

ACID ordering leads to the same

See What Is ACID? on page 7. account balance. If there is an
operation in the set that checks
Big Data for a negative balance and
The volume and variety of charges a fee, then the order in
information collected. Big data which the operations are
is an evolving term that applied absolutely matters.
describes any large amount of
structured, semi-structured, and CRDTs
unstructured data that has the Conflict-free, replicated data-
potential to be mined for infor types are collection data struc
mation. Although big data tures designed to run on
doesnt refer to any specific systems with weak CAP consis
quantity, the term is often used tency, often across multiple data
when speaking about petabytes centers. They leverage commu
and exabytes of data. Big data tativity and monotonicity to
systems facilitate the explora achieve strong eventual guaran
tion and analysis of large data tees on replicated state.
sets. VoltDB is not big data, but
Compared to strongly consis
it does support analytical capa
tent structures, CRDTs offer
bilities using Hadoop, a big data
weaker guarantees, additional
database.
complexity, and can require
CAP additional space. However, they
See What Is CAP? on page 9. remain available for writes dur
ing network partitions that
Commutative Operations would cause strongly consistent
A set of operations are said to systems to stop processing.
be commutative if they can be
applied in any order without Delivery Guarantees
affecting the ending state. See Chapter 5.
For example, a list of account Determinism
credits and debits is considered In data management, a deter
commutative because any ministic operation is one that

35
Dimension Data

will always have the exact same loaded into the long-term data
result given a particular input store.
and state. Determinism is
important in replication. A Exponential Backoff
deterministic operation can be Exponential backoff is a way to
applied to two replicas, assum manage contention during fail
ing the results will match. ure. Often, during failure, many
Determinism is also useful in clients try to reconnect at the
log replay. Performing the same same time, overloading a recov
set of deterministic operations a ering system.
second time will give the same Exponential backoff is a strategy
result. of exponentially increasing the
timeouts between retries on fail
Dimension Data
ure. If an operation fails, wait
Dimension data is infrequently
one second to retry. If that retry
changing data that expands
fails, wait two seconds, then
upon data in fact tables or event
four seconds, etc,. This allows
records.
simple one-off failures to
For example, dimension data recover quickly, but for more-
may include products for sale, complex failures, there will
current customers, and current eventually be a low-enough load
salespeople. The record of a to successfully recover. Often
particular order might reference the growing timeouts are cap
rows from these tables so as not ped at some large number to
to duplicate data. Dimension bound recovery times, such as
data not only saves space, it 16 seconds or 32 seconds.
allows a product to be renamed
and have that rename reflected Fast Data
in all open orders instantly. The processing of streaming
Dimensional schemas also allow data at real-time velocity, ena
easy filtering, grouping, and bling instant analysis, aware
labeling of data. ness, and action. Fast data is
data in motion, streaming into
In data warehousing, a single applications and computing
fact table, a table storing a environments from hundreds of
record of facts or events, com thousands to millions of end
bined with many dimension pointsmobile devices, sensor
tables full of dimension data, is networks, financial transactions,
referred to as a star schema. stock tick feeds, logs, retail sys
tems, telco call routing and
ETL
authorization systems, and
Extract, transform, load is the
more.
traditional sequence by which
data is loaded into a database. Systems and applications
Fast data pipelines may either designed to take advantage of
compress this sequence, or per fast data enable companies to
form analysis on or in response make real-time, per-event deci
to incoming data before it is sions that have direct, real-time

36 | Glossary
 Real-Time Analytics

impact on business interactions See Chapter 6 for a detailed dis

and observations. cussion of idempotence, includ
ing a recipe explaining the use
Fast data operationalizes the
of idempotent processing.
knowledge and insights derived
from big data and enables Metadata
developers to design fast data Metadata is data that describes
applications that make real- other data. Metadata summari
time, per-event decisions. These zes basic information about
decisions may have direct data, which can make finding
impact on business results and working with particular
through streaming analysis of instances of data easier.
interactions and observations,
which enables in-transaction Operational Analytics (another term for
decisions to be made. operational BI)
Operational analytics is the pro
HTAP cess of developing optimal or
Hybrid transaction/analytical realistic recommendations for
processing (HTAP) architec real-time, operational decisions
tures, which enable applications based on insights derived
to analyze live data as it is cre through the application of stat
ated and updated by istical models and analysis
transaction-processing func against existing and/or simula
tions, are now realistic and pos ted future data, and applying
sible. these recommendations in real-
time interactions.
From the Gartner 2014 Magic
Quadrant: they must use the Operational Database
data from transactions, observa Operational database manage
tions, and interactions in real ment systems (also referred to
time for decision processing as as OLTP, or On Line Transac
part of, not separately from, the tion Processing databases) are
transactions. This process is the used to manage dynamic data in
definition of HTAP (for further real time. These types of databa
details, see Hype Cycle for In- ses allow you to do more than
Memory Computing, 2014). simply view archived data.
Source: Gartner, Inc. Analyst: Operational databases allow you
Massimo Pezzini; Hybrid to modify that data (add,
Transaction/Analytical Process change, or delete) in real time.
ing Will Foster Opportunities
for Dramatic Business Innova Real-Time Analytics
tion, January 2014. Real-time analytics is an over
loaded term. Depending on
Idempotence context, real-time means differ
An idempotent operation is an ent things. For example, in
operation that has the same many OLAP use cases, real-time
effect no matter how many can mean minutes or hours; in
times it is applied. fast data use cases, it may mean
milliseconds. In one sense, real-
time implies that analytics can

Glossary | 37
Probabilistic Data Structures

be computed while a human data structure. However, in

waits. That is, answers can be many cases, these inconsisten
computed while a human waits cies are either allowable or can
for a web dashboard or report trigger a broader, slower search
to compute and redraw. on a complete data structure.
This hybrid approach allows
Real-time also may imply that
many of the benefits of using
analytics can be done in time to
probability, and also can ensure
take some immediate action.
correctness of values.
For example, when a user uses
too much of their mobile data Shared Nothing
plan allowance, a real-time ana A shared-nothing architecture
lytics system can notice this and is a distributed computing
trigger a text message to be sent architecture in which each node
to that user. is independent and self-
sufficient and there is no single
Finally, real-time may imply
point of contention across the
that analytics can be computed
system. More specifically, none
in time for a machine to take
of the nodes share memory or
action. This kind of real-time is
disk storage.
popular in fraud detection or
policy enforcement. The analy Streaming Analytics
sis is done between the time a Streaming analytics platforms
credit or debit card is swiped can filter, aggregate, enrich, and
and the transaction is approved. analyze high-throughput data
from multiple disparate live
Probabilistic Data Structures
data sources and in any data
Probabilistic data structures are
format to identify simple and
data structures that have a prob
complex patterns to visualize
abilistic component. In other
business in real time, detect
words, there is a statistically
urgent situations, and automate
bounded probability for cor
immediate actions (definition:
rectness (as in Bloom filters).
Forrester Research).
In many probabilistic data
Streaming operators include:
structures, the access time or
Filter, Aggregate, Geo, Time
storage can be an order of mag
windows, Temporal patterns,
nitude smaller than an equiva
and Enrich.
lent non-probabilistic data
structure. The price for this sav Translytics
ings is the chance that a given Transactions and analytics in
value may be incorrect, or it the same database (source: For
may be impossible to determine rester Research).
the exact shape or size of a given

38 | Glossary
About the Authors
Ryan Betts is one of the VoltDB founding developers and is pres
ently VoltDB CTO. Ryan came to New England to attend WPI. He
graduated with a B.S. in Mathematics and has been part of the Bos
ton tech scene ever since. Ryan has been designing and building dis
tributed systems and high-performance infrastructure software for
almost 20 years. Chances are, if youve used the Internet, some of
your ones and zeros passed through a slice of code he wrote or tes
ted.
John Hugg, founding engineer & Manager of Developer Relations at
VoltDB, specializes in the development of databases, information
management software, and distributed systems. As the first engineer
on the VoltDB product, he worked with the team of academics at
MIT, Yale, and Brown to build H-Store, VoltDBs research prototype.
John also helped build the world-class engineering team at VoltDB
to continue development of the companys open source and com
mercial products. He holds a B.S. in Mathematics and Computer
Science and an M.S. in Computer Science from Tufts University.