An experimental study of online database index selection

algorithms with performance guarantees

Tanzid Sultan

August 2024

Project Supervisors:
Dr. William Umboh
Dr. Junhao Gan
Dr. Renata Borovica-Gajic

A Research Proposal Submitted for the COMP90055 subject at the University of

Melbourne.
Contents
1 Introduction and Background 1
1.1 Database index selection problem . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The need for online index tuning . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Existing methodologies for online index tuning . . . . . . . . . . . . . . 2
1.4 Lack of efforts in establishing performance guarantees . . . . . . . . . . 3
1.5 Limitations of the what-if cost model . . . . . . . . . . . . . . . . . . . . 4

2 Research Objectives 5
2.1 Online index selection using MTS: extending WFIT . . . . . . . . . . . 5
2.2 Dynamic environments: non-stationary and non-stochastic bandits . . . 6
2.3 Comparing WFIT and Bandit Performance . . . . . . . . . . . . . . . . 7

3 Project Plan Outline 8

i
1 Introduction and Background
1.1 Database index selection problem
In today’s era of Big Data, organizations across various sectors - such as finance, health-
care, social media and e-commerce - rely on vast amounts of data to carry out their
daily operations and for strategic planning and decision-making. A Database Manage-
ment System (DBMS) is a software system designed to robustly store and manage such
collections of data. A DBMS also provides an interface that allows end-users to query
and update the data. The speed and efficiency with which a DBMS is able to process
user queries/updates depends critically on the physical design structures (PDS) avail-
able within the system. These are data structures which specify how the data is stored
on disk, retrieved and updated. We will restrict our attention to databases which follow
the relational model [24] in which the data is organized across multiple tables related
to each other through keys.
A PDS on each table primarily consist of heap files, in which the data is un-
sorted, and clustered indexes in which the data is sorted according to a primary key
which uniquely identifies the data records. These PDS do not efficiently support most
queries which are based on columns other than the primary key. This is where sec-
ondary/nonclustered indexes 1 come into play . A database index serves a similar pur-
pose as an index at the end of a book, i.e. it allows us to quickly retrieve the specific
rows we are interested in without performing a full-table scan. Because the main per-
formance bottleneck in large-scale relational databases is the disk I/O, an intelligently
designed set of secondary indexes, also referred to as a configuration, can significantly
reduce the amount of disk I/O needed to process a query resulting in dramatic speedup.

1.2 The need for online index tuning

Most DBMS do not automatically create secondary indexes. Instead, either an expert
database administrator (DBA) needs to manually select and materialize a set of sec-
ondary indexes or a non-expert user can invoke offline index recommendation tools pro-
vided by the DBMS, such as the Database Tuning Advisor in Microsoft SQL Server [1].
Such offline selection/tuning methods require a representative workload to be provided
based on which index recommendations are generated. This approach poses several
challenges, particularly when workload patterns change unpredictably over time due
1
Secondary indexes come in various types, such as B + -tree index, hash index, etc. We will re-
strict our attention to only B + -tree indexes due to their efficiency, versatility and prevalence in most
applications.

1
to varying user demands or seasonal trends, or when the scale of the database changes
due to data growth or increasing number of users. Such dynamic environments would
necessitate periodically invoking the offline tuning process from scratch, each time with
a new representative workload. These offline methods also don’t take into considera-
tion, the overheads associated with generating recommendations and materializing the
recommended indexes. This can often lead to a situation where these overheads can
cause significant delays, outweighing future query processing costs.
Given these challenges with offline tuning, there is a strong motivation for develop-
ing automated online index selection algorithms which can continuously monitor
evolving database workloads (and perhaps also evolving external environmental factors)
and adaptively adjust the index configuration based on observed patterns. More for-
mally, we can frame the algorithm as solving a combinatorial optimization program,
where the objective is to make a sequence of decisions, each decision corresponding to
selecting and materializing a set of secondary indexes (i.e. configuration), such that
the total running time of the entire workload and configuration changes is minimized
over an indefinitely long horizon, subject to some memory constraints. The difficulty
in designing these online algorithms comes from the fact that each decision has to be
made given only knowledge of workloads and environment observed in the past and
without knowledge of the future.
A user can only place their trust on an online algorithm if it provides some form
of robust theoretical (worst-case) performance guarantee. Such a guarantee
would ensure that the algorithm performs well under various scenarios and offer pro-
tection against arbitrary performance degradation. In this project, we will restrict our
attention to online index selection algorithms that satisfy this important criterion.

1.3 Existing methodologies for online index tuning

Substantial efforts have been made to address the need for online index tuning, resulting
in a number of notable algorithms over the last few decades. A key component which
is common across all of these different approaches is the use of a cost model to
estimate query execution time in any hypothetical index configuration2 . This cost
estimation is a necessary step because, at a high level, each of these approaches perform
an exploration of some part of the space of all possible configurations to find the best
possible configuration according to some specified criterion. Beyond the commonality,
these algorithms can be broadly divided into two categories, based on the nature of the
cost model.
2
a hypothetical configuration is one that has not been materialized

2
 In the first category, we have algorithms that rely on the internal cost model3
of the query optimizer, such as [5, 12, 15, 19, 21, 22, 28]. In addition to using this
internal cost model, some of these algorithms, such as [5], are also tightly coupled to
the DBMS query optimizer and are invasive in the sense that they intercept the query
plan at various stages of query optimization and make local plan transformations to
enumerate and explore a large number of alternative hypothetical configurations and
their corresponding estimated costs. Among all approaches in the first category, the
WFIT algorithm of [21], which uses a variant of the work function algorithm from
metrical task systems (MTS) [3], is most notable due to its theoretical performance
guarantee in the form of a bounded competitive ratio. To our knowledge, [15] and
[21] are the only two works which have applied MTS to the online index selection
problem and provided proof of bounded competitive ratio4 , with the latter providing a
substantially tighter bound.
In recent years, there has been a shift towards using learning-based approaches
for index selection such as [2, 7, 8, 9, 16, 18, 17, 23, 26, 29]. This brings us to the
second category of algorithms which use an external cost model, that learns directly
from observed data, instead of relying on the DBMS what-if interface. [2] and [8] take
a reinforcement learning (RL) approach, modeling queries and updates as a Markov
decision process (MDP) with states represented by the index configurations, actions
represented by configuration changes and rewards represented by query execution and
configuration change costs. [2] uses least-squares policy iteration while [8] applies deep
Q-learning to solve the MDP. On the other hand, [16, 17, 18] uses a different kind of RL
approach in the form of stateless Multi-Arm Bandits (MAB) where bandit actions/arms
are represented by configuration changes. Among all the algorithms in the second
category, the MAB particularly stands out due to its theoretical performance guarantee
in the form of a sub-linear regret bound and has been shown in experiments to have
superior performance compared to a state-of-the-art offline tuning tool provided by a
commercial DBMS.

1.4 Lack of efforts in establishing performance guarantees

Surveying the literature on this topic of online database tuning also reveals that a vast
majority of the existing techniques make no direct effort towards providing theoretical
performance guarantees. In the study of online optimization, two very different per-
3
also often referred to as a what-if interface [6]
4
[5] also proves a constant competitive ratio for the special case of a single index, however their
analysis does not extend to the general multi-index case

3
formance metrics are typically used, each of them leading to very different algorithms
and methodologies. On the online learning-based algorithms side, the metric of choice
is the cumulative regret which is defined as the difference between the solution cost of
the algorithm and that of the optimal offline static solution5 . On the other side (which
includes MTS), the metric used is the competitive ratio, which is the worst-case ra-
tio between the solution cost of the algorithm and that of the optimal offline dynamic
solution. Aside from the obvious difference between these two metrics, one being ad-
ditive and the other multiplicative in nature, fundamentally the regret measures the
algorithms ability to learn a static concept while the competitive ratio measure the
ability to learn a dynamic concept [27].
In the realm of database online index selection, the two most notable algorithms
which pay attention to the aforementioned performance metrics are:

• MAB algorithm of [16, 17, 18] with its sublinear regret bound

• MTS-based WFIT algorithm of [21] with its bounded competitive ratio

Both MAB and WFIT have individually demonstrated strong performance over a
range of benchmarking tests [11, 18], however a direct comparison between the two has
not been attempted. Also, despite their good performance, each still possess certain
weak-points that we believe can be overcome by implementing appropriate extensions.
This will be discussed further in the Section 2 and will form the basis for this research
project.

1.5 Limitations of the what-if cost model

In section 1.3, we discussed two categories of online index tuning algorithms, differing
by their use of an internal vs. an external cost model. We also noted that over the
years, there has been a gradual shift towards the second category, particularly learning-
based external cost models. This transition is mainly due to several drawbacks of the
query optimizer’s internal cost model.
The query optimizer cost estimates can be highly error prone, mainly due to inaccu-
rate table statistics and cardinality estimation [14]. This becomes especially apparent
when data distributions are highly skewed, as is often the case in realistic scenarios,
where estimations could be off by orders of magnitude. Such errors can lead to large
discrepancies between estimated and actual costs and adversely effect index tuning
5
an optimal offline static solution is also referred to as an optimal policy in reinforcement learning
literature

4
tools [4]. The use of arbitrary units for query optimizer cost estimates also makes it
difficult, if not impossible, to directly compare these estimates with actual observed
costs. There is also no direct way to obtain an index creation cost estimate using the
what-if interface in some DBMS6 . Moreover, what-if calls tend to be slow in practical
situations and there is significant variability in the running-time.

2 Research Objectives
2.1 Online index selection using MTS: extending WFIT
The WFIT algorithm [21] based on metrical task systems has the distinct property of
having a bounded competitive ratio, which gives us an absolute worst-case theoretical
guarantee for how much the solution quality can degrade. This property makes WFIT
an attractive choice as an online index selection algorithm. Experimental studies in [11]
have compared the performance of several online index selection algorithms, including
WFIT and the algorithm from [5], across both static and dynamic workloads. WFIT
is able to consistently show superior performance across a wide range of metrics, such
as fast convergence under dynamic workloads and low variability in index selection
quality.
However, when it comes to the overhead of generating index recommendations,
WFIT has been shown to be the slowest, due to the overhead being dominated by
the large number of what-if calls made to the query optimizer. Aside from this high
overhead of generating recommendations being a weak-point of WFIT, we would also
like to argue that the over-reliance on the what-if calls is itself a major weak-point.
This can be understood by fact that the query optimizer’s internal cost model can
be highly inaccurate and inefficient (see Section 1.5). The experiments in [11] fail to
expose this weakness as they do not incorporate any datasets which have very high
levels of skew in their data distributions.
This motivates the first goal for our project, which is to extend WFIT by entirely
replacing the use of the internal (what-if) cost model with an external cost model
based on machine-learning, which incrementally learns from observations and is inde-
pendent of the query optimizer. This goal directly aims to remedy the aforementioned
weak-points of WFIT, which includes increasing the accuracy of cost estimation while
significantly reducing the overhead. To achieve this goal, we will execute the following
tasks:
6
e.g. Microsoft SQL Server does not provide the functionality for estimating hypothetical index
creation cost

5
 • Identify and implement an algorithm for the external cost model which has fast,
efficient and robust performance. From our preliminary survey of possible candi-
dates, we believe an online/incremental ridge regression model [20] with carefully
engineered feature inputs might be a good choice.

• Implement the WFIT algorithm and integrate with the external cost model

• Carry out experiments with highly skewed datasets to study and verify the per-
formance benefits of using an external cost model

2.2 Dynamic environments: non-stationary and non-stochastic ban-

dits
The MAB algorithm for online index selection is based on the stochastic bandit assump-
tion, according to which the reward for each arm is drawn i.i.d. from a fixed/stationary
probability distribution. Under certain real-world settings, this assumption is often
violated. For instance, in a multi-tenant cloud environment [10] where compute and
storage resources are shared among multiple independent DBMS instances serving large
number of users, the load on such shared servers may vary greatly depending on envi-
ronmental factors beyond the control of the DBMS. As a result, the amount of resources
available to a DBMS could change unpredictably which can then affect query process-
ing times. This scenario of a dynamic environment can manifest as non-stationarity
of bandit arm reward distributions. The regret bound of the stochastic MAB may not
protect us from arbitrary performance degradation in such a scenario.
The experimental investigations in [16, 17, 18] do not consider such dynamic en-
vironments and steps for extending the stochastic MAB in this setting have not been
explicitly discussed. Therefore our second goal for this project is to design controlled
experiments for simulating a dynamic environment which could expose the weak point
of the stochastic bandit, then consider strategies for extending the MAB index selection
algorithm enablilng it to handle dynamic environments. From our preliminary surveys,
we have identified two suitable directions for achieving this goal.
One possibility is a straight-forward extension of the contextual stochastic bandit,
whereby we directly augment the context vector with additional features which can
track changes in the reward distribution [13, 25]. For instance, in a multi-tenant
database environment, these features could perhaps include information pertaining to
the load that the server is currently experiencing. This contextual information could
then be utilized in the decision-making for super-arm selection.

6
 Another direction, which is not so straight-forward, is to utilise a non-stochastic 7
contextual bandit algorithm such as [30]. The non-stochastic bandit model removes
the assumption of i.i.d. rewards coming from a fixed distributions and allows the
rewards to come from any non-stationary distribution, even from an adversarial pro-
cess. Non-stochastic bandit algorithms vary significantly from deterministic stochastic
bandit approaches (e.g. UCB), they require some randomized elements [13] and im-
plementation is known to be highly non-trivial for problems with large action spaces.
The non-stochastic bandit algorithm [30] has been shown to have a sub-linear regret
bound.

To achieve our second goal, we will be executing the following steps:

• Implement both the simple extension to the stochastic contextual bandit and the
non-stochastic bandit algorithm from [30].

• Design controlled experiments simulating non-stationary reward distributions. A

very simple strategy that we have identified for simulating a non-stationary re-
ward distribution is to add an extra shift value to observed query execution/index
creation costs.

• Run the experiments, evaluate and compare the performance of the stochastic
bandit with the extended approaches and study the impact of non-stationary
rewards.

2.3 Comparing WFIT and Bandit Performance

The WFIT and Bandit algorithms have very different performance guarantees, the for-
mer having bounded competitive ratio and the latter having bounded regret. A unified
theoretical analysis for comparing these two different metrics directly may not be possi-
ble. Hence, in this project, we will not focus on this particular aspect. That being said,
our third and final goal for this project is to carry out a suite of experiments similar
to [16], in order to benchmark the performance of the extended Bandit algorithms, the
extended WFIT algorithm, the built-in tuning advisor tool of Microsoft SQL server
in addition to a no-index baseline. The experimental aspects of goals 1 and 2 will be
covered under goal 3. Detailed logistics regarding the experiments are yet to be fully
determined and the depth of experimentation will depend on amount of time remaining
after completion of the first two goals.
7
the non-stochastic bandit is also sometimes referred to as an adverserial bandit

7
3 Project Plan Outline
Given the remaining duration for the project encompasses weeks 4-12 of the current
semester, we are providing the following week-by-week plan outline for achieving the
tasks in our 3 goals:

• Weeks 4, 5 - Implement external cost model and extended WFIT for goal 1, run
preliminary tests.

• Weeks 6, 7 - Implement extended Bandit for goal 2, run preliminary tests.

• Weeks 8, 9 - Implement experiment suites for goal 3, carry out analysis of results

• Weeks 10, 11, 12 - Prepare thesis report and presentation.

References
[1] Sanjay Agrawal et al. “Database Tuning Advisor for Microsoft SQL Server 2005”.
In: VLDB. Very Large Data Bases Endowment Inc., 2004.
[2] Debabrota Basu et al. “Regularized Cost-Model Oblivious Database Tuning with
Reinforcement Learning”. In: Transactions on Large-Scale Data- and Knowledge-
Centered Systems XXVIII - Volume 9940. Berlin, Heidelberg: Springer-Verlag,
2016, 96–132.
[3] Allan Borodin, Nathan Linial, and Michael E. Saks. “An optimal on-line algo-
rithm for metrical task system”. In: J. ACM 39.4 (1992), 745–763.
[4] Renata Borovica, Ioannis Alagiannis, and Anastasia Ailamaki. “Automated phys-
ical designers: what you see is (not) what you get”. In: Proceedings of the Fifth
International Workshop on Testing Database Systems. DBTest ’12. Scottsdale,
Arizona: Association for Computing Machinery, 2012.
[5] Nicolas Bruno and Surajit Chaudhuri. “An Online Approach to Physical Design
Tuning”. In: 2007 IEEE 23rd International Conference on Data Engineering.
2007, pp. 826–835.
[6] Surajit Chaudhuri and Vivek Narasayya. “AutoAdmin “what-if” index analysis
utility”. In: SIGMOD Rec. 27.2 (1998), 367–378.
[7] Bailu Ding et al. “AI Meets AI: Leveraging Query Executions to Improve In-
dex Recommendations”. In: Proceedings of the 2019 International Conference on
Management of Data. SIGMOD ’19. Amsterdam, Netherlands: Association for
Computing Machinery, 2019, 1241–1258.

8
 [8] Jianling Gao et al. “Automatic index selection with learned cost estimator”. In:
Information Sciences 612 (2022), pp. 706–723.
[9] Benjamin Hilprecht and Carsten Binnig. “Zero-shot cost models for out-of-the-
box learned cost prediction”. In: Proc. VLDB Endow. 15.11 (2022), 2361–2374.
[10] IBM. What is multi-tenant (or multitenancy)? [https://www.ibm.com/topics/
multi-tenant].
[11] Ivo Jimenez et al. “Benchmarking Online Index-Tuning Algorithms”. In: IEEE
Data Eng. Bull. 34.4 (2011), pp. 28–35.
[12] Hai Lan, Zhifeng Bao, and Yuwei Peng. “An Index Advisor Using Deep Rein-
forcement Learning”. In: Proceedings of the 29th ACM International Conference
on Information & Knowledge Management. CIKM ’20. Virtual Event, Ireland:
Association for Computing Machinery, 2020, 2105–2108.
[13] Tor Lattimore and Csaba Szepesvári. Bandit Algorithms. Cambridge University
Press, 2020.
[14] Guy Lohman. “Is Query Optimization a “Solved” Problem?” In: ACM SIGMOD
Blog (2014). [https://wp.sigmod.org/?p=1075].
[15] Tanu Malik et al. “Adaptive Physical Design for Curated Archives”. In: Scien-
tific and Statistical Database Management. Ed. by Marianne Winslett. Berlin,
Heidelberg: Springer Berlin Heidelberg, 2009, pp. 148–166.
[16] R. Malinga Perera et al. “DBA bandits: Self-driving index tuning under ad-hoc,
analytical workloads with safety guarantees”. In: 2021 IEEE 37th International
Conference on Data Engineering (ICDE). Los Alamitos, CA, USA: IEEE Com-
puter Society, 2021, pp. 600–611.
[17] R. Malinga Perera et al. “HMAB: self-driving hierarchy of bandits for integrated
physical database design tuning”. In: Proc. VLDB Endow. 16.2 (2022), 216–229.
[18] R. Malinga Perera et al. “No DBA? No Regret! Multi-Armed Bandits for Index
Tuning of Analytical and HTAP Workloads With Provable Guarantees”. In: IEEE
Transactions on Knowledge amp; Data Engineering 35.12 (2023), pp. 12855–
12872.
[19] Kai-Uwe Sattler, Ingolf Geist, and Eike Schallehn. “QUIET: continuous query-
driven index tuning”. In: Proceedings of the 29th International Conference on
Very Large Data Bases - Volume 29. VLDB ’03. Berlin, Germany: VLDB En-
dowment, 2003, 1129–1132.

9
[20] Rob Schapire and Maryam Bahrani. Online Linear Regression. https://www.
cs . princeton . edu / courses / archive / spring18 / cos511 / scribe _ notes /
0411.pdf. [Lecture Notes for COS 511: Theoretical Machine Learning]. 2018.
[21] Karl Schnaitter and Neoklis Polyzotis. “Semi-automatic index tuning: keeping
DBAs in the loop”. In: Proc. VLDB Endow. 5.5 (2012), 478–489.
[22] Karl Schnaitter et al. “On-Line Index Selection for Shifting Workloads”. In:
2007 IEEE 23rd International Conference on Data Engineering Workshop. 2007,
pp. 459–468.
[23] Tarique Siddiqui et al. “DISTILL: low-overhead data-driven techniques for filter-
ing and costing indexes for scalable index tuning”. In: Proc. VLDB Endow. 15.10
(2022), 2019–2031.
[24] Abraham Silberschatz, Henry F. Korth, and S. Sudarshan. Database System Con-
cepts. 6th ed. New York: McGraw-Hill, 2010. Chap. 2.
[25] Aleksandrs Slivkins. Introduction to Multi-Armed Bandits. 2024. arXiv: 1904 .
07272 [cs.LG].
[26] Ji Sun and Guoliang Li. “An end-to-end learning-based cost estimator”. In: Proc.
VLDB Endow. 13.3 (2019), 307–319.
[27] Adam Wierman. “A tale of two metrics: competitive ratio and regret”. In: (2014).
[https://rigorandrelevance.wordpress.com/2014/10/13/a-tale-of-two-
metrics-competitive-ratio-and-regret/].
[28] Wentao Wu et al. “Budget-aware Index Tuning with Reinforcement Learning”.
In: Proceedings of the 2022 International Conference on Management of Data.
SIGMOD ’22. Philadelphia, PA, USA: Association for Computing Machinery,
2022, 1528–1541.
[29] Tao Yu et al. “Refactoring Index Tuning Process with Benefit Estimation”. In:
Proc. VLDB Endow. 17.7 (2024), 1528–1541.
[30] Lukas Zierahn et al. “Nonstochastic Contextual Combinatorial Bandits”. In:
Proceedings of The 26th International Conference on Artificial Intelligence and
Statistics. Ed. by Francisco Ruiz, Jennifer Dy, and Jan-Willem van de Meent.
Vol. 206. Proceedings of Machine Learning Research. PMLR, 2023, pp. 8771–
8813.

10