Query optimization is expert work: read the plan, find the stage that hurts, and pick the rewrite, index or hint that fixes it. We turned that expertise into an LLM-based query advisor for Apache Pinot, a distributed real-time analytics database. It reads the query, the execution plan, the runtime statistics and the table configuration, and recommends what to change.

The model alone was not good enough, and this talk shows how we fixed that. We wrote down what a human expert checks as patterns the model applies. Deterministic gates reject advice that does not match the table. Each rewrite runs against the live cluster before a user sees it. An evaluation suite measures every prompt change. You will leave with a recipe for putting an LLM in front of a database without letting it hurt you when it is wrong.

Optimizing distributed SQL is very hard, even for experts. Apache Pinot provides very low-level information to optimize queries, but as one of the query experts at StarTree, I can say LLMs are significantly more efficient than humans at doing so