Training a 4B model to produce 81% faster query plans than Postgres

Language models can now create database query plans that outperform traditional systems.

A project trained a 4-billion parameter language model, Qwen, to optimize database query plans, achieving an 81% speedup over standard PostgreSQL query plans. The training involved agentic reinforcement learning, where Qwen generated candidate query plans and received feedback based on execution time compared to PostgreSQL's default. This process iteratively improved Qwen's ability to generate faster plans. The experiment successfully reduced query latency by 44.7% across 113 join-heavy queries, with the model initially being unable to generate plans for most of them. The setup included a custom reinforcement learning environment designed to minimize noise from factors like Linux page caching and utilized a measurement rig across separate machines for training and execution.

This experiment demonstrates a novel application of language models beyond natural language processing, directly impacting database performance and query optimization. By leveraging reinforcement learning, the 4B parameter model Qwen was fine-tuned to discover and suggest more efficient query execution strategies than a mature system like PostgreSQL. This suggests a potential shift in how database performance is approached, moving towards AI-driven optimization rather than solely relying on heuristic-based algorithms embedded in database engines.

The success of training a relatively small model like Qwen for this specific task implies that specialized AI agents could become common components in database management systems, offering significant performance gains for complex queries. Organizations managing large datasets and complex analytical workloads may see substantial improvements in query response times and resource utilization, potentially reducing infrastructure costs. Future developments will likely focus on integrating such AI optimizers directly into database software or as standalone performance tuning tools.