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

Rohan Bansal has introduced Qorl, an innovative approach to database query optimization that leverages a 4-billion parameter machine learning model. By replacing traditional heuristic-based query planners with a learned model, the project demonstrates a significant performance improvement, achieving query plans that execute 81% faster than those generated by the standard PostgreSQL optimizer. The research explores how deep learning can be integrated into database internals to handle complex query structures more efficiently than static cost-based models. Bansal details the training methodology, the challenges of mapping SQL structures to model inputs, and the performance benchmarks observed during testing. This development highlights a growing trend of applying generative AI and neural networks to core infrastructure components, potentially transforming how relational databases manage execution strategies and resource allocation in high-load environments.