Overview
Production Readiness
0.7
Novelty Score
0.65
Cost Impact Score
0.7
Citation Count
0
Why It Matters For Business
E3-Rewrite can cut query costs by reducing runtime for heavy analytical queries while keeping results correct. That directly lowers compute bills and improves interactive analytics latency for complex queries.
Summary TLDR
E3-Rewrite fine-tunes large language models with execution-aware prompts and reinforcement learning to produce SQL rewrites that are executable, semantically equivalent, and faster. Key ingredients: prepend EXPLAIN plan text to prompts, retrieve similar successful rewrites, and train with a two-stage RL curriculum that first enforces correctness then optimizes latency.
Problem Statement
Rule-based SQL rewriting is brittle and hard to extend. LLMs can express richer rewrites but often produce non-executable or non-equivalent SQL and can hurt performance. We need a learnable system that reliably generates executable, equivalent, and faster SQL.
Main Contribution
E3-Rewrite: end-to-end LLM-based SQL rewriting that optimizes for executability, equivalence, and efficiency without hand-coded rules.
Execution-hint prompts: prepend linearized EXPLAIN / EXPLAIN ANALYZE plans so the model sees operator choices and bottlenecks.
Two-stage RL (GRPO) curriculum: stage 1 enforces executability and equivalence; stage 2 adds a performance reward to reduce latency.
Hybrid demonstration retrieval: retrieve past high-quality rewrite pairs by structural (AST) and semantic similarity and update the pool online.
Key Findings
Big average latency reduction on TPC-H.
High correctness while improving many queries.
Reinforcement learning is essential for equivalence.
Execution-plan hints materially reduce latency.
Results
TPC-H average latency (Original)
TPC-H p90 latency
Improved queries (>=10% speedup)
Equivalence ratio after rewriting
Ablation: equivalence without RL
Ablation: effect of removing execution plan hints on avg latency
Who Should Care
What To Try In 7 Days
Run EXPLAIN for a few slow, important queries and prepend the plan text to an LLM prompt to see whether simple plan-aware edits help.
Collect a small demo pool of original+good-rewrite pairs and try retrieval-based prompting to reuse patterns.
Add a post-check that compares outputs between original and rewritten queries on sample DB instances before deploying rewrites.
Agent Features
Memory
- dynamic demonstration pool (online updates of rewrite pairs)
Planning
- curriculum: correctness → performance
Tool Use
- uses EXPLAIN / EXPLAIN ANALYZE as external tool
- uses QEDsolver for formal equivalence checks
Frameworks
- GRPO
Architectures
- GRPO
- two-stage curriculum learning
Optimization Features
Training Optimization
- GRPO
- curriculum-guided two-stage training
- rehearsal of Stage 1 examples during Stage 2
Inference Optimization
- prepend linearized execution plans to prompts
- retrieve top-k hybrid demonstrations (structure + semantics)
Reproducibility
Data Urls
- TPC-H (public)
- IMDB / JOB (public)
- DSB (public)
Data Available
Open Source Status
- partial
Risks & Boundaries
Limitations
- Experiments run on PostgreSQL v14; optimizer behavior may differ across DBMSs.
- Equivalence checking relies on QEDsolver + LLM fallbacks; solver timeouts force heuristic decisions.
- Estimated cost from EXPLAIN (optimizer estimates) can diverge from real runtime, causing imperfect perf signals.
- Closed-source LLM baselines (GPT-4o) limit head-to-head reproducibility for some comparisons.
When Not To Use
- When formal proof of semantic equivalence is mandatory and solver support is required for all constructs.
- On DBMSs that do not expose comparable EXPLAIN plans or have very different optimizer IDs.
- For trivial, short queries where rule-based fixes are simpler and cheaper.
Failure Modes
- LLM generates syntactically valid but semantically different SQL; fallback checks may be inconclusive.
- Optimizer cost estimates mislead the performance reward, causing rewrites that appear better to EXPLAIN but worsen real latency.
- Demonstration pool contamination: adding low-quality rewrites could bias retrieval and outputs.
Core Entities
Models
- Qwen (Qwen 14B)
- Qwen (Qwen 32B)
- LLaMA4
- GPT-4o
Metrics
- Query Latency
- Equivalence Ratio
- Improved Queries (>=10% speedup)
Datasets
- TPC-H
- IMDB (JOB)
- DSB
Benchmarks
- TPC-H
- IMDB (JOB)
- DSB