A <scp>ndromeda</scp> : Debugging Database Performance Issues with Retrieval-Augmented Large Language Models

Debugging performance issues in a database management system (DBMS) is tedious and challenging, even for experienced database administrators (DBAs). Thus, with the rapid advancement of large language models (LLMs), developing an LLM-powered co-pilot to assist or even replace DBAs by automatically diagnosing issues and generating recommendations for resolution presents a promising direction. However, directly prompting LLMs for DBMS performance debugging often yields either generic or irrelevant responses, as LLMs may lack both domain knowledge in performance debugging and a deep understanding of DBMS internals. In this paper, we introduce Andromeda, a retrieval-augmented, LLM-powered system for automatic DBMS performance debugging. Andromeda enables users to pose natural language questions about various performance issues and provides context-aware and actionable recommendations for resolution. To achieve this, Andromeda first retrieves relevant evidence from multiple sources, including historical questions, troubleshooting manuals, DBMS telemetry data, and execution logs. It then leverages effective training strategies to adapt an open-source LLM, empowering it to diagnose and resolve issues based on the retrieved evidence. We have developed a web application for Andromeda and demonstrated its effectiveness in debugging real-world DBMS performance issues.