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

1 INTRODUCTION

• Existing tuners treat configuration spaces as box-constrained and learn documented knob dependencies through trial and error, wasting evaluations on invalid configurations.Hundreds of knobs and broad vendor ranges make tuning costly.Source: §1 · p. 1
• CATune encodes deterministic ordering constraints in the search domain and restricts Bayesian optimization to a constraint-consistent subspace.Unlike methods that learn feasibility boundaries from sampled violations, it enforces constraints during both exploitation and exploration.Source: §1 · p. 2
• Topology-aware sampling respects the dependency graph during generation, producing feasible configurations without rejection or repair overhead.Post-hoc rejection or repair can incur substantial overhead when multiple dependencies apply.Source: §1 · p. 2
• CATune improves final tuning quality and convergence speed across workloads, surrogate models, BO implementations, and search-space regimes.The introduction also reports that correct ordering constraints prevent startup failures and improve reliability, while hallucinated constraints can harm optimization.Source: §1 · p. 2
• CATune’s precision-first LLM-assisted pipeline discovers and refines knob dependencies while mitigating hallucinated constraints.The paper reports that direct LLM prompting can produce unsupported or incorrectly oriented dependencies, including cases that degraded tuning efficiency.Source: §1 · p. 2
• The introduction also compares constraint-handling strategies and analyzes constraint use, hallucination risks, and system-level effects.These analyses accompany the system’s experiments on tuning performance and convergence.Source: §1 · p. 2

2 MOTIVATION

• In preliminary PostgreSQL TPC-C experiments with 46 knobs, approximately 35% of tuning iterations led to DBMS crashes, while dependency violations also produced rejected or semantically meaningless executions.Such wasted iterations increase operational cost and slow model convergence.Source: §2 · p. 2
• Approximately 21% higher throughput within the first 100 iterations was observed with constraints; the min_wal_size ≤ max_wal_size constraint eliminates nearly 50% of the nominal search region and reduces invalid samples by 19.5%.Figure 1 presents preliminary TPC-C tuning with and without constraints and illustrates feasible and infeasible regions for this ordering constraint.Source: Figure 1 · p. 3, §2 · p. 3
• Although tuning knowledge is documented in manuals and guidelines, explicit inter-knob dependencies remain largely unmodeled and are left for tuners to learn from sampled observations.The section notes that surrogate models are not guaranteed to faithfully capture structural knob relationships.Source: §2 · p. 3
• Dependency extraction is challenging because constraints appear in varied linguistic forms, and errors may prune feasible configurations or bias optimization.The proposed response combines schema-constrained LLM parsing, symbolic normalization, rule-based patterns, and reliability checks.Source: §2 · p. 3
• Because BO and RL frameworks target box-constrained black-box search, deterministic inter-knob dependencies require explicit integration into optimization.The section frames direct enforcement of documented dependencies as a requirement for effective DBMS tuning.Source: §2 · p. 3

3 PROBLEM FORMULATION

• The classical formulation represents a configuration as a vector of knobs drawn from continuous, integer, or categorical domains.The nominal space is the Cartesian product of the individual knob domains; performance is an unknown function evaluated by executing the workload.Source: §3 · p. 3
• Documented pairwise ordering constraints define an admissible configuration space, excluding structurally invalid settings before performance evaluation.The constraints are known from documentation and can be verified without running the DBMS.Source: §3 · p. 3
• CATune’s overview combines manual-based ordering-constraint extraction with a Bayesian optimization engine that uses those dependencies during exploration and exploitation.Source: Figure 2 · p. 4

3.1 Computational Complexity

• The constraint-aware tuning problem remains NP-hard under a black-box objective, even when dependencies are ordering constraints.The proof sketch notes that an unconstrained binary instance may require evaluating all 2^n configurations.Source: §3.1 · p. 4
• Checking whether pairwise ordering constraints are consistent is tractable in polynomial time, unlike optimizing over the feasible region.The constraints can be checked through cycle detection or a shortest-path feasibility check.Source: §3.1 · p. 4

Workflow

• CATune takes the target DBMS and manuals, tuning knobs and ranges, objective, and workload, then searches until the user’s resource budget is exhausted.It extracts dependencies into machine-readable constraints before optimization.Source: §4 · p. 4

Precision-first Extraction

• Its precision-first extraction pipeline combines schema-constrained LLM parsing, symbolic normalization, rule templates, and verification guardrails to produce machine-readable constraints.The records can include a knob pair, relation, optional activation condition, and evidence span.Source: §4 · p. 4

Exploration phase

• During exploration, CATune builds a topology graph, harmonizes units, normalizes knobs by component, and samples feasible configurations without rejection.Samples are mapped back to their original value domains.Source: §4 · pp. 4–5

Exploitation phase

• During exploitation, structural constraints restrict the acquisition function’s feasible region so proposed trials avoid infeasible configurations.Source: §4 · p. 5

5.1 Exploitation Phase

• CATune restricts acquisition maximization to dependency-feasible candidates without modifying the surrogate model.Candidate selection is limited to the intersection of the local neighborhood and the feasible configuration space.Source: §5.1 · p. 5

5.2 Exploration Phase

• The exploration phase applies a five-step procedure to maintain constraint consistency throughout the search.Figure 3 illustrates the end-to-end workflow.Source: §5.2 · p. 5, Figure 3 · p. 6

5.2.1 Topology Graph.

• Ordering constraints form a directed graph whose edge directions encode which knob must be assigned first.Cycle detection identifies inconsistent constraints; otherwise, the graph is a DAG defining a partial order among dependent knobs.Source: §5.2.1 · p. 5

5.2.2 Search Space Normalization.

• Compatible knobs are converted to canonical units and normalized using shared intervals for each connected component, preserving ordering consistency.The shared interval spans the component's normalized lower and upper bounds; for example, work_mem ≤ shared_buffers uses [4, 8192] MB rather than independently normalized domains.Source: §5.2.2 · p. 6

5.2.3 Topology-aware Sampling.

• Topology-aware sampling assigns roots first, then samples each dependent knob within parent-constrained ranges, making every configuration feasible by construction.This eliminates rejection and gives sampling cost linear growth with the number of knobs; inverse normalization restores native DBMS units.Source: Figure 3 · p. 6, §5.2.3 · p. 7, §5.2.2 · p. 7

6.2 Precision-First Extraction Pipeline

• The precision-first pipeline retrieves candidate passages, extracts schema-constrained records with evidence spans, verifies support and operand direction, then normalizes relations and resolves duplicates or conflicts.Uncertain records are retained only when two judge calls agree on the same canonical relation; otherwise, the pipeline abstains.Source: §6 · p. 7, §6.2 · p. 8

6.1 Extraction Scope and Inputs

• The optimizer enforces verified pairwise numerical orderings, while activation conditions, resource bounds, and multi-knob dependencies are retained but not translated into topology constraints.Extraction requires a version-specific official documentation corpus, target knob list, and predefined relation schema; knob selection itself is outside scope.Source: §6.1 · p. 7

6.3 Comparison with Manual-reading Systems

• Unlike manual-reading systems that extract suggested values for individual knobs, CATune extracts inter-knob dependencies and converts supported pairwise orderings into structural constraints.Records represent the knob pair, relation direction, optional condition, supporting evidence, and confidence.Source: §6.3 · p. 8

7.2 Efficiency Evaluation

• Under default ranges, CATune reached baseline SMAC optimum 12.5× faster on TPC-C; with suggested ranges, throughput improved 63.37% on TPC-C.Under default ranges, throughput improved 33.39% on TPC-C and latency decreased 12.55% on TPC-H; with suggested ranges, CATune reached the baseline optimum 5.41× faster on TPC-C and 3.57× faster on TPC-H, and reduced TPC-H latency by 10.73%.Source: Figure 4 · p. 9, Figure 5 · p. 9, §7.2 · p. 9

7.1 Experimental Setup

• Experiments used PostgreSQL v13 and MySQL v8.0, 45 and 50 performance-relevant knobs, 10 ordering dependencies, and 200 tuning iterations.Four runs with different random seeds were averaged; objectives were throughput for TPC-C and average latency for TPC-H.Source: §7.1 · p. 9

7.3 Generalizing across Optimizers

• With GP-BO, CATune reached baseline optimum 8.33× faster and improved TPC-C throughput by 28.71%; with GPTuner, it was 5.26× faster and improved throughput by 4.81%.The authors report consistent gains across random-forest and Gaussian-process surrogates and across standard and coarse-to-fine BO frameworks.Source: Figure 6 · p. 10, Figure 7 · p. 10, §7.3 · p. 10

7.4 Scalability Study

• The scalability study examines whether CATune’s benefits persist as database size increases, but its detailed results are provided only in the technical report.The study is omitted from the main paper due to space limitations.Source: §7.4 · p. 10

7.5 Ablation Study

• Against rejection-based rule enforcement, topology-aware sampling achieved 18.8% higher throughput and reached the same performance level 2.19× faster.The comparison isolates topology-aware sampling: it generates feasible configurations directly rather than repeatedly rejecting infeasible ones.Source: Figure 8 · p. 10, §7.5 · p. 10

8.1 Operational Impact of Constraints

• Nine of 200 sampled configurations failed to start, all violating superuser_reserved_connections < max_connections; two injected hallucinated constraints reduced the achievable improvement by 32.07% relative to baseline.Without hallucinations, CATune reached baseline optimum 4.55× faster and achieved a 59.15% final-performance improvement; it achieved 134.29% higher final performance and converged 9.09× faster than the hallucinated setting.Source: §8.1 · p. 11

8.2 Efficiency of Different Sampling Strategies

• Topology-aware sampling returned 100 feasible configurations from 100 generated candidates, while rejection sampling generated 15,992 candidates and took 6.841 seconds to obtain 100 feasible samples.Unconstrained sampling produced 99 invalid configurations out of 100; repair also required modifying 99 out of 100 samples.Source: Table 2 · p. 11, §8.2 · p. 11

8.3 Impact of Different Constraint Employment Strategies

• CATune reached baseline optimum 3.93× faster on TPC-C and 3.57× faster on TPC-H; MultiObj achieved 2.3× and 1.74× speedups, respectively.After 200 iterations, CATune improved TPC-C throughput by 53.52% over baseline versus 38.26% for MultiObj, and reduced TPC-H latency by 10.73% versus 3.43%.Source: §8.3 · p. 11, Figure 10 · p. 12

9 RELATED WORK

• Database BO systems typically assume box-constrained spaces, while classical constrained BO models unknown or partially observed constraints and may sample infeasible configurations to learn their boundaries.ResTune handles workload-dependent service-level constraints through post-hoc runtime measurements.Source: §9 · p. 12
• CATune encodes static, system-defined ordering dependencies directly into the domain, so sampled configurations are valid by construction rather than checked after execution.The paper distinguishes these dependencies from workload-dependent constraints and notes that violations may cause instability or startup failures.Source: §9 · p. 12
• Unlike GPTuner’s range refinement and parameter prioritization, CATune explicitly enforces structural validity among knobs at the configuration-space level.The paper presents the approaches as complementary: range refinement biases exploration, whereas CATune encodes deterministic ordering constraints.Source: §9 · p. 12

10 LIMITATIONS

• The current optimizer uses ordering and resource-bounded constraints; incorporating richer extracted constraint types remains future work.The extraction pipeline also identifies activation conditions, multi-knob dependencies, and performance recommendations.Source: §10 · p. 12
• Ordering constraints are uncommon in some DBMSs, comprising only 2.1% in ClickHouse, which limits how many additional restrictions CATune can impose there.The paper positions CATune as complementary to existing tuners where fewer supported constraints are available.Source: §10 · p. 12
• Integrating CATune with low-dimensional projections needs specialized designs because many-to-one knob mappings may not preserve constraints on the original knobs.Developing constraint-preserving projections remains future work.Source: §10 · p. 12

• CATune incorporates deterministic knob dependencies directly into Bayesian optimization.Source: §11 · p. 12
• By eliminating infeasible configurations by construction, CATune improves tuning reliability and efficiency.Source: §11 · p. 12
• The results demonstrate the value of explicitly modeling structural dependencies in DBMS configuration spaces.Source: §11 · p. 12