Pith. sign in

REVIEW 2 major objections 4 minor 54 references

AutoIndexer: A Reinforcement Learning-Enhanced Index Advisor Towards Scaling Workloads

T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AutoIndexer claims that workload compression and state-wise RL action masking jointly solve the scaling problem for index advisors, making deep-RL index selection practical for workloads of over one thousand queries.

desk verdict The paper's core compression ILP doesn't actually compress; the idea is promising but the central formulation is unsound as written. read the letter →

arxiv 2507.23084 v1 pith:FP47VDJ7 submitted 2025-07-30 cs.DB cs.AI

Taiyi Wang , Eiko Yoneki This is my paper
classification cs.DBcs.AI
keywords indexselectionreinforcementlearningworkloadcompressionactionmaskingOLAPdatabasetuningscalabilityqueryrewrite
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

AutoIndexer seeks to establish that deep reinforcement learning can be made to scale for index selection by tackling the two bottlenecks—workload size and action-space explosion—together rather than separately. The paper claims that a column-centric compression stage, which prunes queries and columns through a weighted-bipartite-graph integer program, reduces the search space enough for a state-wise RL agent with learned action masking to find high-quality index sets in a fraction of the time. If the claim holds, database administrators could apply RL-based index tuning to workloads of more than one thousand queries, with up to 95% end-to-end runtime reduction versus no indexes, about 20% better workload cost savings than prior RL advisors, and more than 50% lower tuning time.

What carries the argument

The central mechanism is a pipeline of two coupled pruners. First, the compressor builds a weighted bipartite graph between queries and indexable columns, with edge weights $w_{i,j}$ combining the estimated speedup $\Delta\text{Cost}(q_i,c_j)$ of indexing column $c_j$ for query $q_i$ and a synergy term that adds the speedup of similar queries, then solves a small integer linear program to keep a representative query subset. Second, the index-selection stage augments a discrete on-policy actor-critic with a selector network $G_\theta$ that, per state, produces a binary mask over the candidate pool; a Thompson sampler turns the logits into a pruned action set, and the selector is trained by comparing TD errors from masked and unmasked actions. The preprocessor wraps this with SMT-verified query rewrites and a workload state that concatenates plan embeddings, tokenized-query embeddings, and an index-configuration bitmap.

What would settle it

Instrument the compression stage: feed it a small workload with known positive per-query speedups and run the ILP exactly as written (maximizing $\sum w_{ij}x_iy_j$ subject only to $\sum x_iy_j \ge \text{Cover}(E)$). If the solver returns the full workload—as the objective suggests it must—then the compression benefit reported in the paper depends on an additional constraint that is never spelled out.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that workload compression and RL action-space optimization reinforce each other: compressing at column granularity keeps the information the RL policy needs while shrinking the candidate set, and the RL agent's learned action masking then concentrates exploration on high-value indexes. The reported result is that this co-design matches or exceeds the index quality of uncompressed tuning while cutting selection time from hours to tens of seconds, and it scales to 1,120-query workloads that prior RL advisors cannot handle without stalling.

Load-bearing premise

The compression ILP works only if some mechanism actually discards queries and columns; as written, maximizing a non-negative weighted sum under a lower-bound coverage constraint would keep everything, so the method's reported pruning must rest on an unstated cardinality or sparsity constraint whose soundness is not demonstrated.

Editorial extensions

If this is right

  • Index selection for workloads of more than one thousand queries can complete in minutes, not hours, bringing RL-based advisors into practical tuning windows.
  • The compressor alone, when ported to heuristic or other RL tuners, still reduces selection time, showing the compression benefit is not an artifact of the RL module.
  • Replacing the AutoIndexer compressor with Isum raises both selection time and workload cost, indicating that column-level correlation-preserving merges matter more than compression size.
  • Each module earns its place: removing the compressor raises costs by up to 7x, removing the RL model raises TPC-H cost from 5‰ to 25‰, and removing the workload model costs 9% on JOB.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same weighted-bipartite-graph machinery could summarize workloads for other physical-design tasks, such as partitioning or materialized view selection, where query-column dependency structure is equally load-bearing.
  • The learned workload embeddings, finetuned to be invariant to literal values, could plausibly transfer to query plan prediction or join-order learning; the paper does not evaluate transfer beyond index selection.
  • The vectorized Index-Gym and compressed state suggest a practical continual-learning setup: fine-tune the trained selector on a sliding window of compressed queries as workloads drift, rather than retraining from scratch, which addresses the paper's stated limitation about rapidly changing workloads.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. AutoIndexer proposes a three-stage pipeline for index selection on large analytical workloads: (1) a workload-compression stage that removes duplicate or similar queries, ranks and trims columns, and then selects a representative query subset using a weighted bipartite graph and an integer program; (2) a preprocessing stage with SMT-guided query rewriting, candidate-index enumeration, and a learned workload model; and (3) an RL stage built on a state-wise actor-critic with Thompson-sampled action masking. The paper reports up to 95% end-to-end query-time reduction versus no-index baselines, roughly 20% average cost-savings improvement over RL baselines, and more than 50% tuning-time reduction on TPC-H, TPC-DS, and JOB. The central scaling mechanism is the compression stage, which is claimed to reduce workload size before RL training begins.

Significance. If the compression formulation were sound, the combination of column-centric workload reduction with RL action masking would be a practically interesting contribution, and the modular ablation design in Figures 10-13 is a sensible way to isolate component contributions. The paper also deserves credit for reporting training-time limitations and for comparing against external baselines such as SWIRL, Lan et al., Extend, AutoAdmin, and Isum rather than only self-comparisons. However, the central compression ILP is unsound as written, and the scalability claims in Figures 8-13 and Tables 4-6 rest directly on that stage. The internal inconsistency in Section 4.3.2 and Table 3 further weakens the training-efficiency narrative. As a result, the current manuscript does not support its main claims.

major comments (2)
  1. [§4.3.2, Table 3] The narrative claim that Lan et al. and SWIRL 'both struggled beyond N=100, often requiring 10-15+ h or failing to converge' is directly contradicted by Table 3, which reports Lan at 3.3 h for N=500 and 5.3 h for N=1120, and SWIRL at 11.1 h for N=500; only SWIRL at N=1120 exceeds 15 h. The text should be reconciled with the table, or the table corrected; as written, the training-efficiency comparison is internally inconsistent.
  2. [§4.3.2, robustness paragraph] The robustness statement 'When randomly replacing 20% to 50%-80% of seen templates/queries, performance drops by 13%-21% respectively, but still outperforms Extend and AutoIndexer by 30% on solution quality' is internally confused: it names AutoIndexer itself as a baseline, and the range '20% to 50%-80%' is ambiguous. This makes the claimed quantitative robustness result difficult to interpret and should be rewritten with a clear experimental protocol and a well-defined baseline set.
minor comments (4)
  1. [§4.3.1] The sentence 'AutoIndexer achieves 40‰ while AutoAdmin remains at 300‰' mixes permille and percentage notation; please state clearly whether these are percentages or permille values to avoid a 10× ambiguity.
  2. [Table 3 and §3.2.2] Table 3 uses maximum index width W_max=3, while the rest of the paper, including the candidate-enumeration description, uses W_max=5; please state whether this difference is intentional and whether the scalability claims depend on it.
  3. [§3.1.2, hyperparameter configuration] The text says compression parameters are selected 'automatically' and 'eliminate the need for manual parameter tuning,' but then describes an empirically derived 85% column-ranking threshold, a shallow grid search over delta-cost values and balance coefficients, and a tunable Cover(E); these statements should be reconciled.
  4. [§6 and artifact statement] Despite the 'artifact announcement' in Section 6, no code or data link appears in the text; if artifacts are available, please include a URL and a reproducibility statement.

Circularity Check

1 steps flagged · score 6.0 of 10

Compression ILP is a no-op: with non-negative weights its optimum is the full workload, so the compressed set equals its input by construction.

  1. other [Section 3.1.2, Phase 3, Step 3 (ILP-based subset selection)]
    "We then solve a small Integer Linear Program whose objective is to maximise the total collected edge weight while guaranteeing that every critical column is still covered: Maximise ∑_{(i,j)∈E} w_{i,j} x_i y_j subject to ∑_{(i,j)∈E} x_i y_j ≥ Cover(E), x_i∈{0,1}, y_j∈{0,1} ∀i,j. Cover(E) is a tunable lower bound on the number of distinct query–column pairs that must remain. The decision variables y_j merely ensure coverage; our final output is the set of queries with x_i = 1."

    All components of w_{i,j} are non-negative (ΔCost speed-up, query frequency f_i, and Jaccard similarity), so every additional retained query-column pair adds non-negative objective value. The only constraint is a lower bound on retained pairs, which the all-ones assignment satisfies; no cardinality or sparsity term appears in the objective. Therefore the ILP optimum includes every query and column, and the 'compressed workload' Q-hat is the Phase-2 workload itself by construction. The paper's claim that 'the ILP decide[s] how many queries to retain' and the compression-driven scaling results (Figures 8-13, Tables 4-6) do not follow from the stated equations; the compression stage is a no-op unless an unstated cardinality constraint is added.

full rationale

The central end-to-end comparison is not circular: AutoIndexer's cost-savings claims are tested against external baselines (SWIRL, Lan, Extend, AutoAdmin), and the self-citations [34,41] used to contrast the state-wise masking mechanism are not load-bearing evidence for those comparisons. The one by-construction degeneracy is in the compression ILP (Section 3.1.2, Phase 3, Step 3). With non-negative edge weights and only a lower-bound coverage constraint, the stated optimization is maximized by retaining the full Phase-2 workload, so the ILP output equals its input. The paper explicitly says 'the ILP decide[s] how many queries to retain', but the equations provide no incentive to discard any query. Because workload compression is the first stage and is invoked to explain scalability (Figures 8-13, Tables 4-6), this is a load-bearing step whose claimed compression effect reduces to identity as written. If the implementation includes an unstated cardinality or sparsity constraint, the paper must state it; otherwise this is a partial circularity: the compression 'prediction' is equivalent to no compression by construction. The RL and workload-model modules still have independent empirical content, so the score is partial rather than total.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central method rests on two unproven domain assumptions: that optimizer cost estimates are faithful and that column-level compression preserves index quality. The paper also hides several tuned parameters such as column threshold, similarity threshold, lambda values, and Cover(E), which appear benchmark-specific. Because no code or data is shipped, it is impossible to tell which assumptions are load-bearing in the implementation.

free parameters (7)
  • Column retention threshold = 85%
    Top 85% of columns retained based on Pareto analysis in Section 3.1.2 Phase 2; controls action-space size and is tuned to the benchmarks.
  • Synergy weighting lambda (compression) = not specified
    Lambda in the edge weight formula in Section 3.1.2 Phase 3 adjusts synergy contributions; set by shallow grid search.
  • Query similarity threshold = not specified
    Jaccard threshold above which queries are merged in Section 3.1.2 Phase 1; no numeric value given.
  • Coverage lower bound Cover(E) = not specified
    Tunable lower bound on retained query-column pairs in the ILP in Section 3.1.2 Phase 3; the paper does not state how it is set or whether it bounds the query set.
  • Maximum index width W_max = 5
    User-specified bound on index columns in Section 3.2.2; limits candidate permutations.
  • Sparsity weight lambda (selector loss) = not specified
    Sparsity regularizer in the selector loss in Equation (2) in Section 3.3.3; no value given.
  • Storage budgets and training steps = 0.25-5 GB; k*5000 steps
    Benchmark-specific budgets and vectorized environment steps in Sections 4.1 and 3.3.2; these are experiment settings, not fitted physics.
assumptions (5)
  • domain assumption Optimizer cost estimates from PostgreSQL and HypoPG are reliable proxies for actual query execution time.
    Used throughout Phase 2 and 3 for ColBenefit and reward; if what-if estimates misrank indexes, the RL signal is wrong.
  • domain assumption Compressing the workload by removing queries and columns preserves the relative quality of candidate index sets.
    The entire method depends on the compressed workload guiding RL to near-optimal indexes; this is asserted in O4 and Section 3.1 but not proven.
  • domain assumption Template-generated queries from TPC-H and TPC-DS qgen and the JOB benchmark represent real analytical workloads.
    Evaluation claims generalization to diverse workloads from these benchmarks in Section 4.1.
  • standard math SMT equivalence checking of rewrite rules is sound for the query fragments used.
    The rewriter relies on Z3 unsat results to certify equivalence in Section 3.2.1; this assumes the U-expression encoding and the SMT solver are correct.
  • domain assumption The RL policy trained on compressed workloads transfers to the original workload distribution.
    Training uses 80 to 1120 query subsets and evaluation uses similar template distributions; transfer to unseen templates is only partially tested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AutoIndexer: A Reinforcement Learning-Enhanced Index Advisor Towards Scaling Workloads." pith.science (2026). https://pith.science/paper/FP47VDJ7

@misc{pith2026250723084,
  author       = {Pith},
  title        = {Pith review of: AutoIndexer: A Reinforcement Learning-Enhanced Index Advisor Towards Scaling Workloads},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FP47VDJ7}},
  note         = {Machine review of arXiv:2507.23084}
}
read the original abstract

Efficiently selecting indexes is fundamental to database performance optimization, particularly for systems handling large-scale analytical workloads. While deep reinforcement learning (DRL) has shown promise in automating index selection through its ability to learn from experience, few works address how these RL-based index advisors can adapt to scaling workloads due to exponentially growing action spaces and heavy trial and error. To address these challenges, we introduce AutoIndexer, a framework that combines workload compression, query optimization, and specialized RL models to scale index selection effectively. By operating on compressed workloads, AutoIndexer substantially lowers search complexity without sacrificing much index quality. Extensive evaluations show that it reduces end-to-end query execution time by up to 95% versus non-indexed baselines. On average, it outperforms state-of-the-art RL-based index advisors by approximately 20% in workload cost savings while cutting tuning time by over 50%. These results affirm AutoIndexer's practicality for large and diverse workloads.

Figures

Figures reproduced from arXiv: 2507.23084 by the authors.

Figure 1
Figure 1. Showcase of the critical role of effective index ad [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Motivation and challenges in scaling-aware index selection (Workload size denotes the number of queries in a single [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Performance comparisons showing the impact of [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: System pipeline of AutoIndexer of index selection for large workloads, workload compression tech￾niques are crucial for scalable index tuning. 3.1.1 Targets and Motivation. Workload compression addresses the scalability challenges in index tuning by reducing large quer…
Figure 5
Figure 5. Figure 5: Query selection via weighted bipartite graph [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Pre-processing Procedures for State-Wise RL [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: State-wise reinforcement learning framework in [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of comparative approaches vs AutoIndexer for a Scaling Join Order Benchmark workload on PostgreSQL. Chart: cost relative to processing w/o indexes; Table: selection runtime. 0.25 GB and 5 GB. The classical approach Extend requires over two hours at larger bu…
Figure 9
Figure 9. Figure 9: End2End workload runtime improvements among [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 12
Figure 12. Figure 12: (a) Training behavior comparison between our proposed [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 11
Figure 11. Figure 11: Sole effects on the compressor, combined with [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 46 canonical work pages

  1. [1]

    Gordon, and Bohan Zhang

    Dana Van Aken, Andrew Pavlo, Geoffrey J. Gordon, and Bohan Zhang. 2017. Automatic Database Management System Tuning Through Large-Scale Machine Learning. In Proceedings of the 2017 ACM SIGMOD International Conference on Management of Data. 1009–1024

    work page 2017

  2. [2]

    Charalampos P Andriotis and Konstantinos G Papakonstantinou. 2019. Managing engineering systems with large state and action spaces through deep reinforce- ment learning. Reliability Engineering & System Safety 191 (2019), 106483

    work page 2019

  3. [3]

    Matteo Brucato, Tarique Siddiqui, Wentao Wu, Vivek Narasayya, and Surajit Chaudhuri. 2024. Wred: Workload Reduction for Scalable Index Tuning. Proceed- ings of the ACM on Management of Data 2, 1 (2024), 1–26

    work page 2024

  4. [4]

    Nicolas Bruno and Surajit Chaudhuri. 2005. Automatic physical database tuning: A relaxation-based approach. In Proceedings of the 2005 ACM SIGMOD interna- tional conference on Management of data . 227–238

    work page 2005

  5. [5]

    Surajit Chaudhuri and Vivek R Narasayya. 1997. An efficient, cost-driven index selection tool for Microsoft SQL server. InVLDB, Vol. 97. San Francisco, 146–155

    work page 1997

  6. [6]

    Jin Chen, Guanyu Ye, Yan Zhao, Shuncheng Liu, Liwei Deng, Xu Chen, Rui Zhou, and Kai Zheng. 2022. Efficient Join Order Selection Learning with Graph-based Representation. In Proceedings of the 28th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) . 97–107

    work page 2022

  7. [7]

    Shumo Chu, Brendan Murphy, Jared Roesch, Alvin Cheung, and Dan Suciu. 2018. Axiomatic foundations and algorithms for deciding semantic equivalences of SQL queries. arXiv preprint arXiv:1802.02229 (2018)

    arXiv 2018

  8. [8]

    HypoPG Contributors. 2024. HypoPG: Hypothetical Indexes for PostgreSQL. https://github.com/HypoPG/hypopg. Accessed: 2024-01-01

    work page 2024

Show all 54 references
  1. [9]

    Brian F Cooper, Adam Silberstein, Erwin Tam, Raghu Ramakrishnan, and Russell Sears. 2010. Benchmarking cloud serving systems with YCSB. In Proceedings of the 1st ACM symposium on Cloud computing . 143–154

    2010

  2. [10]

    Leonardo De Moura and Nikolaj Bjørner. 2008. Z3: An efficient SMT solver. In International conference on Tools and Algorithms for the Construction and Analysis of Systems. Springer, 337–340

    2008

  3. [11]

    Shaleen Deep, Anja Gruenheid, Paraschos Koutris, Jeffrey Naughton, and Stratis Viglas. 2020. Comprehensive and efficient workload compression. arXiv preprint arXiv:2011.05549 (2020)

    2020 arXiv

  4. [12]

    Yue Han, Guoliang Li, Haitao Yuan, and Ji Sun. 2021. AutoView: An Autonomous Materialized View Management System with Deep Reinforcement Learning. In Proceedings of the 37th IEEE International Conference on Data Engineering (ICDE) . 289–300

    2021

  5. [13]

    Stratos Idreos, Martin L Kersten, and Stefan Manegold. 2009. Self-organizing tuple reconstruction in column-stores. In Proceedings of the 2009 ACM SIGMOD International Conference on Management of data . 297–308

    2009

  6. [14]

    Stratos Idreos, Stefan Manegold, Harumi Kuno, and Goetz Graefe. 2011. Merging what’s cracked, cracking what’s merged: adaptive indexing in main-memory column-stores. Proceedings of the VLDB Endowment 4, 9 (2011), 586–597

    2011

  7. [15]

    Jan Kossmann, Alexander Kastius, and Rainer Schlosser. 2022. SWIRL: Selection of Workload-aware Indexes using Reinforcement Learning.. In EDBT, Vol. 2. 155–2

    2022

  8. [16]

    Hai Lan, Zhifeng Bao, and Yuwei Peng. 2020. An index advisor using deep reinforcement learning. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management . 2105–2108

    2020

  9. [17]

    Guoliang Li, Xuanhe Zhou, Shifu Li, and Bo Gao. 2019. QTune: A Query-Aware Database Tuning System with Deep Reinforcement Learning. Proc. VLDB En- dowment 12, 12 (2019), 2118–2130

    2019

  10. [18]

    Elmore, and Sanjay Krishnan

    Xi Liang, Aaron J. Elmore, and Sanjay Krishnan. 2019. Opportunistic View Materialization with Deep Reinforcement Learning. CoRR abs/1903.01363 (2019)

    2019 arXiv

  11. [19]

    Vincent Y Lum and Huei Ling. 1971. An optimization problem on the selection of secondary keys. In Proceedings of the 1971 26th annual conference . 349–356

    1971

  12. [20]

    Sultan J Majeed and Marcus Hutter. 2021. Exact reduction of huge action spaces in general reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 8874–8883

    2021

  13. [21]

    Ryan Marcus, Parimarjan Negi, Hongzi Mao, Nesime Tatbul, Mohammad Al- izadeh, and Tim Kraska. 2021. Bao: Making Learned Query Optimization Practical. In Proceedings of the 2021 ACM SIGMOD International Conference on Management of Data. 1275–1288

    2021

  14. [22]

    Ryan Marcus, Parimarjan Negi, Hongzi Mao, Chi Zhang, Mohammad Alizadeh, Tim Kraska, Olga Papaemmanouil, and Nesime Tatbul. 2019. Neo: A Learned Query Optimizer. Proc. VLDB Endowment 12, 11 (2019), 1705–1718

    2019

  15. [23]

    Ryan Marcus and Olga Papaemmanouil. 2018. Deep Reinforcement Learning for Join Order Enumeration. In Proceedings of the 1st Intl. Workshop on Exploiting AI Techniques for Data Management (aiDM@SIGMOD). 1–4

    2018

  16. [24]

    Suphakit Niwattanakul, Jatsada Singthongchai, Ekkachai Naenudorn, and Su- pachanun Wanapu. 2013. Using of Jaccard coefficient for keywords similarity. In Proceedings of the international multiconference of engineers and computer scientists, Vol. 1. 380–384

    2013

  17. [25]

    Daniel J Russo, Benjamin Van Roy, Abbas Kazerouni, Ian Osband, Zheng Wen, et al. 2018. A tutorial on thompson sampling. Foundations and Trends ® in Machine Learning 11, 1 (2018), 1–96

    2018

  18. [26]

    Zahra Sadri, Le Gruenwald, and Eleazar Lead. 2020. DRLindex: deep reinforce- ment learning index advisor for a cluster database. In Proceedings of the 24th Symposium on International Database Engineering & Applications . 1–8

    2020

  19. [27]

    Rainer Schlosser, Jan Kossmann, and Martin Boissier. 2019. Efficient scalable multi-attribute index selection using recursive strategies. In 2019 IEEE 35th International Conference on Data Engineering (ICDE) . IEEE, 1238–1249

    2019

  20. [28]

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov

  21. [29]

    Ankur Sharma, Felix Martin Schuhknecht, and Jens Dittrich. 2018. The case for automatic database administration using deep reinforcement learning. arXiv preprint arXiv:1801.05643 (2018)

    2018 arXiv

  22. [30]

    Tarique Siddiqui, Saehan Jo, Wentao Wu, Chi Wang, Vivek Narasayya, and Surajit Chaudhuri. 2022. ISUM: Efficiently compressing large and complex workloads for scalable index tuning. In Proceedings of the 2022 International Conference on Management of Data. 660–673

    2022

  23. [31]

    Tarique Siddiqui and Wentao Wu. 2024. ML-Powered Index Tuning: An Overview of Recent Progress and Open Challenges. ACM SIGMOD Record 52, 4 (2024), 19–30

    2024

  24. [32]

    John M Stephens and Meikel Poess. 2009. Converting TPC-H Query Templates to Use DSQGEN for Easy Extensibility. In Performance Evaluation and Benchmark- ing: First TPC Technology Conference, TPCTC 2009, Lyon, France, August 24-28, 2009, Revised Selected Papers 1 . Springer, 99–115

    2009

  25. [33]

    Lohman, Volker Markl, and Mokhtar Kandil

    Michael Stillger, Guy M. Lohman, Volker Markl, and Mokhtar Kandil. 2001. LEO: DB2’s Learning Optimizer. In Proceedings of the 27th International Conference on Very Large Data Bases (VLDB) . 19–28

    2001

  26. [34]

    Hao Sun and Taiyi Wang. 2022. Toward Causal-Aware RL: State-Wise Action- Refined Temporal Difference. arXiv preprint arXiv:2201.00354 (2022)

    2022 arXiv

  27. [35]

    Richard S Sutton and Andrew G Barto. 2018. Reinforcement learning: An intro- duction. MIT press

    2018

  28. [36]

    Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulao, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al. 2024. Gymnasium: A standard interface for reinforcement learning environments. arXiv preprint arXiv:2407.17032 (2024)

    2024 arXiv

  29. [37]

    Immanuel Trummer, Junxiong Wang, Deepak Maram, Samuel Moseley, Saehan Jo, and Joseph Antonakakis. 2019. SkinnerDB: Regret-Bounded Query Evaluation via Reinforcement Learning. In Proceedings of the 2019 ACM SIGMOD International Conference on Management of Data . 1153–1170

    2019

  30. [38]

    Gary Valentin, Michael Zuliani, Daniel C Zilio, Guy Lohman, and Alan Skelley

  31. [39]

    Hado Van Hasselt, Arthur Guez, and David Silver. 2016. Deep reinforcement learning with double q-learning. InProceedings of the AAAI conference on artificial intelligence, Vol. 30

    2016

  32. [40]

    Jiangqing Wang, Yinglun Hu, Chong Sun, Huili Zhang, and Chongwei Ruan

  33. [41]

    Taiyi Wang and Eiko Yoneki. 2024. IA2: Leveraging Instance-Aware Index Advisor with Reinforcement Learning for Diverse Workloads. In Proceedings of the 4th Workshop on Machine Learning and Systems . 10–17

    2024

  34. [42]

    Zhaoguo Wang, Zhou Zhou, Yicun Yang, Haoran Ding, Gansen Hu, Ding Ding, Chuzhe Tang, Haibo Chen, and Jinyang Li. 2022. Wetune: Automatic discovery and verification of query rewrite rules. In Proceedings of the 2022 International Conference on Management of Data . 94–107

    2022

  35. [43]

    Kyu-Young Whang. 1987. Index selection in relational databases. In Foundations of Data Organization. Springer, 487–500

    1987

  36. [44]

    Yu Yan, Shun Yao, Hongzhi Wang, and Meng Gao. 2021. Index selection for NoSQL database with deep reinforcement learning. Information Sciences 561 (2021), 20–30

    2021

  37. [45]

    Zongheng Yang, Wei-Lin Chiang, Sifei Luan, Gautam Mittal, Michael Luo, and Ion Stoica. 2022. Balsa: Learning a Query Optimizer Without Expert Demonstrations. In Proceedings of the 2022 ACM SIGMOD International Conference on Management of Data. 931–944

    2022

  38. [46]

    Xiang Yu, Chengliang Chai, Guoliang Li, and Jiabin Liu. 2022. Cost-based or Learning-based? A Hybrid Query Optimizer for Query Plan Selection. Proc. VLDB Endowment 15, 13 (2022), 3924–3936

    2022

  39. [47]

    Xiang Yu, Guoliang Li, Chengliang Chai, and Nan Tang. 2020. Reinforcement Learning with Tree-LSTM for Join Order Selection. In Proceedings of the 36th IEEE International Conference on Data Engineering (ICDE) . 1297–1308

    2020

  40. [48]

    Ji Zhang, Yu Liu, Ke Zhou, and Guoliang Li. 2019. An End-to-End Automatic Cloud Database Tuning System Using Deep Reinforcement Learning. In Proceed- ings of the 2019 ACM SIGMOD International Conference on Management of Data . 415–432

    2019

  41. [49]

    Wei Zhou, Chen Lin, Xuanhe Zhou, and Guoliang Li. 2024. Breaking It Down: An In-Depth Study of Index Advisors. Proceedings of the VLDB Endowment 17, 10 (2024), 2405–2418. Taiyi Wang and Eiko Yoneki

    2024

  42. [50]

    Xuanhe Zhou, Guoliang Li, Chengliang Chai, and Jianhua Feng. 2021. A learned query rewrite system using monte carlo tree search. Proceedings of the VLDB Endowment 15, 1 (2021), 46–58

    2021

  43. [51]

    Rong Zhu, Wei Chen, Bolin Ding, Xingguang Chen, Andreas Pfadler, Ziniu Wu, and Jingren Zhou. 2023. Lero: A Learning-to-Rank Query Optimizer. Proc. VLDB Endowment 16, 6 (2023), 1466–1479

    2023

  44. [2000]

    In Proceedings of 16th International Conference on Data Engineering (Cat

    DB2 advisor: An optimizer smart enough to recommend its own indexes. In Proceedings of 16th International Conference on Data Engineering (Cat. No. 00CB37073). IEEE, 101–110

  45. [2017]

    arXiv preprint arXiv:1707.06347 (2017)

    Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017)

    2017 arXiv

  46. [2024]

    In Proceedings of the 21st IEEE International Conference on e-Business Engineering (ICEBE)

    Joint Recommendation of Materialized Views and Indexes Based on Rein- forcement Learning. In Proceedings of the 21st IEEE International Conference on e-Business Engineering (ICEBE). 212–217

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.