Explaining Database Performance Regressions Using a Large Language Model

Revealing and explaining performance regressions in relational database workloads remain time-consuming tasks for database professionals, especially in distributed database systems. In this paper we use a large language model (LLM), specifically OpenAI’s o4-mini, to identify and explain regressions across different versions of the same database system. We focus on duration metrics available in modern database query histories, without any additional tracing. Experiments were conducted using semantic metric descriptions and structured metrics values as context for the LLM. The model was able to attribute differences to particular phases of execution or metric categories, and to recognize cases where differences in performance were caused by noise. While LLM-generated explanations were often informative, recommended practical actions had limited value. This work demonstrates the potential of LLMs as a new tool in analysis of database systems performance.