How do Java deadlocks happen, and how do you prevent them?

During an incident, I preserve dumps and request symptoms before restarting, then restore capacity without creating a restart loop. I map each held lock back to the code path and define one ordering rule that every path follows. Verification runs the conflicting operations repeatedly under load and confirms progress, bounded wait time, and unchanged invariants. Monitoring includes blocked-thread trends and operation latency, while static analysis or focused tests guard the ordering rule where practical.