Reproduce “Plan-Structured Deep Neural Network Models for Query Performance Prediction” on OpenResearch

About this paper

Ryan Marcus, Olga Papaemmanouil

A Plan-Structured Deep Neural Network (PS-DNN) was developed at Brandeis University to accurately predict database query execution times, eliminating the need for human-engineered features. This model achieved up to a 9% improvement in prediction accuracy over existing state-of-the-art methods on standard benchmarks.

Key insights

• Deep learning, when integrated with a carefully tailored architecture that respects query plan structure, can completely automate Query Performance Prediction, thereby eliminating human-engineered features.
• Designing neural networks that specifically account for properties like branch isolation, heterogeneous nodes, and shared operator behavior is crucial for effective application of deep learning to database problems.
• The PS-DNN provides significantly more accurate and robust latency predictions, particularly for complex and long-running queries where traditional, heuristic-based methods often underperform.

Summary by alphaXiv.