DOT: Dynamic Knob Selection and Online Sampling for Automated Database Tuning

Yifan Wang; Debabrota Basu; Pierre Bourhis; Romain Rouvoy; Patrick Royer

cs.DBMar 16, 2026

DOT: Dynamic Knob Selection and Online Sampling for Automated Database Tuning

Authors: Yifan Wang, Debabrota Basu, Pierre Bourhis, Romain Rouvoy, Patrick Royer

Organizations: Orange / Inria / Univ. Lille, France France · Inria / Univ. Lille / CNRS, France France · CNRS / Univ. Lille / Inria, France France · Univ. Lille / CNRS / Inria, France France ·

Abstract

Database Management Systems (DBMS) are crucial for efficient data management and access control, but their administration remains challenging for Database Administrators (DBAs). Tuning, in particular, is known to be difficult. Modern systems have many tuning parameters, but only a subset significantly impacts performance. Focusing on these influential parameters reduces the search space and optimizes performance. Current methods rely on costly warm-up phases and human expertise to identify important tuning parameters. In this paper, we present DOT, a dynamic knob selection and online sampling DBMS tuning algorithm. DOT uses Recursive Feature Elimination with Cross-Validation (RFECV) to prune low-importance tuning parameters and a Likelihood Ratio Test (LRT) strategy to balance exploration and exploitation. For parameter search, DOT uses a Bayesian Optimization (BO) algorithm to optimize configurations on-the-fly, eliminating the need for warm-up phases or prior knowledge (although existing knowledge can be incorporated). Experiments show that DOT achieves matching or outperforming performance compared to state-of-the-art tuners while substantially reducing tuning overhead.

Figures & tables

Benchmark	Type	Mean	SD	CV (%)
TPC-H	OLAP	80.13 sec.	0.42	0.52
TPC-C	OLTP	79.33 TPS	4.95	6.24
Sysbench	OLTP	608.63 TPS	26.53	4.36

Table 1. Variability of performance measurements

Benchmark	Iterations			Time (hours)
Benchmark	Mean	SD	CV (%)	Mean	SD	CV (%)
TPC-H	57.0	41.3	72.5	1.21	0.9	72.9
TPC-C	215.6	122.2	56.7	6.9	4.4	64.3
Sysbench	118.4	40.3	34.1	3.4	1.2	34.6

Table 2. Convergence variability of the BO tuner: iterations & time to reach near-optimal performance across benchmarks

Workload	Sensitivity	P & B	Lasso	CART	Experience	LLM
Sysbench (TPS)	1411.1	939.1	1469.04	1588.7	1294.36	1411.1
TPC-H (Exec. s)	46.22	74.35	46.02	45.76	45.72	46.35

Table 3. Performance tuned with TOP-5 knobs selected by different methods

Figure 1. Variability of CART knob selection across varying sample sizes & sampling strategies

Samples	Rank (Max TPS)	Rank (tuning time [min])	Sampling
	Test stat: $7.32$	Test stat: $1.32$	time (min)
	p-value: $6.2E-2$	p-value: $7.2E-1$
520	$1.8$ (1,708.0 $\pm$ 29.1)	$2.8$ (216.2 $\pm$ 32.6)	866,7
300	$2.0$ (1,684.3 $\pm$ 48.9)	$2.0$ (160.8 $\pm$ 79.5)	500,0
200	$2.4$ (1,690.8 $\pm$ 65.2)	$2.8$ (214.9 $\pm$ 116.9)	333.3
100	$3.8$ (1,602.3 $\pm$ 59.5)	$2.4$ (181.3 $\pm$ 36.1)	166.7

Table 4. Sysbench tuning with CART on knob sets from varying sample sizes (statistically significant at

p<

5E-2)

	Sysbench	TPC-C	TPC-H	EXP
Sysbench	—	$0.317\pm 0.035$	$0.325\pm 0.033$	$0.258\pm 0.016$
TPC-C	$0.317\pm 0.035$	—	$0.472\pm 0.041$	$0.243\pm 0.029$
TPC-H	$0.325\pm 0.033$	$0.472\pm 0.041$	—	$0.409\pm 0.024$
EXP	$0.258\pm 0.016$	$0.243\pm 0.029$	$0.409\pm 0.024$	—

Table 5. Jaccard index between knob sets.

	Rank (Max TPS)	Rank (tuning time [min])
	Test stat: 12.7, p-value: 2.7E-2	Test stat: 17.8, p-value: 3E-3
Sysbench	$2.4$ (1,708.0 $\pm$ 29.1)	$2.4$ (216.2 $\pm$ 32.6)
TPC-C	$4.0$ (1,653.2 $\pm$ 27.2)	$3.0$ (218.4 $\pm$ 65.3)
TPC-H	$4.0$ (1,638.8 $\pm$ 35.2)	$2.8$ (188.3 $\pm$ 57.9)
EXP	$1.4$ (1,725.2 $\pm$ 31.1)	$2.6$ (217.8 $\pm$ 42.7)
Shared	$3.2$ (1,677.1 $\pm$ 35.2)	$5.8$ (447.3 $\pm$ 26.2)

Table 6. Sysbench tuning with knob sets for different workloads ranked with CART (statistically significant at

p<

5E-2)

	Sysbench		TPC-C		TPC-H
	Rank (Max TPS) Test stat: 13.3 p-value: 1E-2	Rank (tuning time [mins]) Test stat: 9.1 p-value: 5.8E-2	Rank (Max TPS) Test stat: 12.0 p-value: 1.7E-2	Rank (tuning time [mins]) Test stat: 5.1 p-value: 2.8E-1	Rank (Exec (s)) Test stat: 6.1 p-value: 1.9E-1	Rank (tuning time [mins]) Test stat: 11.8 p-value: 1.9E-2
Top 5 knobs	4.6 (1,486.8 ± 159.0)	3.0 (311.5 ± 194.2)	5.0 (120.5 ± 22.1)	2.4 (151.7 ± 161.1)	3.6 (47.7 ± 0.3)	1.6 (46.1 ± 54.5)
Top 10 knobs	4.2 (1,674.9 ± 34.4)	2.6 (272.0 ± 122.8)	3.2 (338.5 ± 13.8)	2.0 (153.8 ± 186.7)	3.8 (47.7 ± 0.5)	2.0 (32.7 ± 14.6)
Top 20 knobs	2.0 (1,725.2 ± 31.1)	2.0 (217.8 ± 42.7)	1.8 (354.2 ± 14.8)	3.4 (411.9 ± 264.9)	3.2 (47.8 ± 1.4)	3.2 (73.0 ± 53.2)
Top 30 knobs	2.0 (1,728.4 ± 46.3)	2.6 (267.9 ± 136.3)	2.4 (347.6 ± 25.8)	3.2 (261.0 ± 287.1)	2.8 (47.8 ± 2.0)	3.6 (189.6 ± 286.5)
All 52 knobs	2.2 (1,720.0 ± 33.0)	4.8 (732.6 ± 266.7)	2.6 (346.0 ± 7.5)	4.0 (560.4 ± 327.7)	1.6 (46.5 ± 0.6)	4.6 (334.1 ± 175.5)

Table 7. Performance, tuning time, and Friedman-test ranks for different numbers of knobs (statistically significant at p < 5e-2)

Figure 2. DOT overview

Figure 3. MAPE of Partial vs. Full Benchmark Results

Figure 4. MAPE vs. Benchmark-Query Coverage

Figure 5. Performance vs. tuning-time of different algorithms on Sysbench , TPC-C, TPC-H with knob importance knowledge

Figure 6. Performance vs. tuning-time of different algorithms on Sysbench , TPC-C, TPC-H without knob importance knowledge

	Sysbench		TPC-C		TPC-H
	Rank (Max TPS) Test stat: 37.4 p-value: 2.2E-05	Rank (tuning time [mins]) Test stat: 32.2 p-value: 1.9E-04	Rank (Max TPS) Test stat: 31.4 p-value: 2.5e-04	Rank (tuning time [mins]) Test stat: 37.9 p-value: 1.7E-05	Rank (exec [s]) Test stat: 18.382 p-value: 3.1e-02	Rank (tuning time [mins]) Test stat: 38.4 p-value: 1.5E-05
DOT +EXP	2.6 (1728.3 ± 28.3)	2.0 (57.7 ± 28.2)	1.2 (377.7 ± 16.9)	2.4 (121.6 ± 72.1)	4.4 (46.6 ± 2.2)	3.4 (57.6 ± 54.3)
DOT +RAN	3.0 (1724.7 ± 50.6)	4.2 (171.1 ± 121.2)	2.0 (370.8 ± 13.6)	3.8 (187.3 ± 107.5)	2.6 (45.5 ± 1.3)	2.0 (33.9 ± 63.9)
LLM knowledge based	8.0 (1489.1 ± 57.5)	3.4 (129.3 ± 99.1)	9.2 (264.9 ± 10.1)	5.4 (288.0 ± 207.9)	6.6 (47.0 ± 0.4)	2.0 (21.1 ± 10.0)
BO+CART	3.6 (1708.0 ± 29.1)	10.0 (1080.3 ± 32.6)	5.0 (356.1 ± 9.4)	9.8 (960.4 ± 57.6)	3.6 (46.4 ± 0.5)	10.0 (1027.4 ± 106.6)
CDBTune cold+EXP	8.6 (1358.1 ± 152.9)	3.2 (218.6 ± 310.1)	8.2 (295.2 ± 51.7)	1.2 (33.5 ± 40.9)	-	-

Table 8. Performance, tuning time, and Friedman-test ranks for different methods (statistically significant at

p<

5E-2)

Figure 7. DOT behavior on Sysbench workload with and without knob importance prior

Figure 8. Tuning Performance of DOT on periodic workloads composed of TPC-C and Sysbench

Figure 9. Max TPS Across Configurations and Workloads

Figure 10. Tuning Performance of DOT on Sysbench for various configurations without knob importance knowledge

Figure 11. Tuning Performance of DOT on Sysbench for various

k_{0}

without knob importance knowledge

Figure 12. Tuning Performance of DOT on TPC-C for various

T_{\text{cut}}

budget with knob importance knowledge

Figure 13. Bayesian optimization & RFECV overhead

Database Systems

See all 44 papers

WAter: A Workload-Adaptive Knob Tuning System based on Workload Compression
Mar 28, 2026Yibo Wang, Jiale Lao, Chen Zhang +3Database Systems Relational Databases
Purdue University, West Lafayette, Indiana, USA · Cornell University, Ithaca, New York, USA · Sichuan University, Chengdu, China
CATune: Structural Constraint-Aware Bayesian Optimization for DBMS Configuration Tuning
Oct 7, 2026Fangping Lan, Qi Zhang, Eduard DragutDatabase Systems Bayesian Optimization
Temple University Philadelphia, PA, USA
A Case for Agentic Tuning: From Documentation to Action in PostgreSQL
May 19, 2026Hongyu Lin, Mingyu Li, Weichen Zhang +4Large Language Model-Guided Optimization LLM Agent Skill Learning
Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · Key Laboratory of System Software (Chinese Academy of Sciences) +3

Mar 28, 2026J/K move · Enter open · S save

WAter: A Workload-Adaptive Knob Tuning System based on Workload Compression

Yibo Wang, Jiale Lao, Chen Zhang, Cehua Yang+2

Purdue University, West Lafayette, Indiana, USA · Cornell University, Ithaca, New York, USA · Sichuan University, Chengdu, China

Database Systems Relational Databases

Open paper AskPDF arXiv

Selecting appropriate values for the configurable parameters of Database Management Systems (DBMS) to improve performance is a significant challenge. Recent machine learning (ML)-based tuning systems have shown strong potential, but their practical adoption is often limited by the high tuning cost. This cost arises from two main factors: (1) the system needs to evaluate a large number of configurations to identify a satisfactory one, and (2) for each configuration, the system must execute the entire target workload on the DBMS, which is both time-consuming. Existing studies have primarily addressed the first factor by improving sample efficiency, that is, by reducing the number of configurations evaluated. However, the second factor, improving runtime efficiency by reducing the time required for each evaluation, has received limited attention and remains an underexplored direction. We develop WAter, a runtime-efficient and workload-adaptive tuning system that finds near-optimal configurations at a fraction of the tuning cost compared with state-of-the-art methods. We divide the tuning process into multiple time slices and evaluate only a small subset of queries from the workload in each slice. Different subsets are evaluated across slices, and a runtime profile is used to dynamically identify more representative subsets for evaluation in subsequent slices. At the end of each time slice, the most promising configurations are evaluated on the original workload to measure their actual performance. Evaluations demonstrate that WAter identifies the best-performing configurations with up to 73.5% less tuning time and achieves up to 16.2% higher performance than the best-performing alternative.

Abstract

Database Systems

Explore similar work

WAter: A Workload-Adaptive Knob Tuning System based on Workload Compression

CATune: Structural Constraint-Aware Bayesian Optimization for DBMS Configuration Tuning

A Case for Agentic Tuning: From Documentation to Action in PostgreSQL