Show HN: I replaced database deadlock victim selection with an LLM — here's the data

Every mainstream database uses fixed rules for deadlock victim selection.

MySQL kills the one with the fewest locks. CockroachDB kills the youngest. PostgreSQL kills the one that closed the cycle. Not one of them can explain "why you" — the killed transaction just knows it got rolled back, no reason given.

I wanted to see what happens if an LLM makes that decision instead.

What I Built

DDA is a Sidecar process. It sits outside the database, polls lock state, builds a wait-for graph, runs DFS cycle detection, and — only at the "choose a victim" step — calls an LLM (deepseek-v4-pro via Anthropic SDK). The other four steps (SQL execution, lock parsing, graph construction, rollback) are pure deterministic Python code.