Show HN: I replaced database deadlock victim selection with an LLM — here's the data
A developer has created a sidecar process called DDA that replaces traditional database deadlock victim selection with an LLM. DDA polls lock states, builds a wait-for graph, and then queries an LLM, specifically DeepSeek V4-Pro via the Anthropic SDK, to choose which transaction to roll back. While LLM latency is a significant trade-off compared to microsecond-level fixed rules, DDA successfully resolved all tested deadlock scenarios and demonstrated more nuanced reasoning than fixed rules by considering lock types and admitting when transactions were symmetric. AI
IMPACT Demonstrates LLMs can offer more nuanced decision-making than fixed rules in specialized technical domains.