LLMs for database tuning

The performance of database systems depends on various tuning choices. The optimal choices depend on workload, data, and hardware properties. The search space for database tuning is large, and reliable cost models (predicting performance for specific tuning choices) are often unavailable. In this line of work, we show how to leverage large language models (LLMs) to make database tuning faster and better.

Our DB-BERT system reads database manuals, blog entries, and Stackoverflow discussions about database tuning to find hints about how to set specific parameters. It then selects optimal combinations of hints via reinforcement learning, basing rewards on performance measurements. λ-Tune leverages more advanced LLMs to generate entire database configuration scripts, based on a description of hardware, workload, and data properties provided to the LLM.

Tuning and designing database system well requires realistic benchmarks. However, real industry workloads cannot be shared due to privacy considerations. The SQLBarber system leverages high-level workload statistics to generate matching SQL benchmarks using LLMs automatically. In another line of work, we leverage LLMs to predict correlation in data from column names (thereby informing query optimization), or apply LLMs to analyze query plans and trace back performance regressions.