HERO

HERO: Hint-based Efficient and Reliable Query Optimizer

A learned query optimization system that addresses three key challenges in hint-based query optimization: reliable hint recommendation, efficient hint exploration, and fast inference.

Key Contributions

• Ensemble of context-aware models for reliable hint suggestion, replacing traditional NN approaches that fail under reliability constraints
• Graph storage architecture enabling fast inference and hint exploration
• Budget-controlled training with parameterized local search to navigate exponential search space efficiently
• Achieves optimization capability close to best achievable with coarse-grained hints on standard benchmarks

Technologies

Python, PyTorch, PostgreSQL, openGauss, Machine Learning

Impact

40× faster inference compared to existing hint-based approaches while maintaining reliability guarantees.

Links

• Paper (arXiv)
• GitHub Repository
• Benchmark Platform