Can LLMs Replace the DBA? Evaluating Ontology-Driven Prompting for Relational Database Tuning Tasks

Abstract

Relational database tuning traditionally relies on specialized heuristics and expert intuition, making the process difficult to scale, audit, and reproduce. While recent research has explored autonomic and AI-based systems to assist Database Administrators, the emergence of Large Language Models introduces a new paradigm for automated tuning. This paper investigates the extent to which LLMs can augment or replace traditional DBA expertise in complex, knowledge-dependent tuning tasks. We evaluate tuning recommendations generated by two distinct LLMs across three model variants each, comparing zero-shot performance against ontology-driven prompting. Using a PostgreSQL environment with a TPC-H-like workload, we measure the power, throughput, and storage footprint of the recommended structures. Our results indicate that expert knowledge—whether from an experienced DBA or a formal ontology—is not merely a source of domain vocabulary. Ultimately, the ontology acts as a critical precision mechanism that codifies the ’rules of thumb’ inherent to expert tuning, effectively grounding the LLM’s generative capacity and optimizing the trade-off between performance and storage footprint.