Graceful shutdown in Node backends only works when the process stops accepting work before it drains in-flight requests. Prefer closing the listen socket first, reject new connections with a clear readiness fail, then wait on outstanding handlers up to a hard deadline before exit. Drain background timers and open clients in the same window so a deploy does not cut mid-write.
Ops prove-out after wiring shutdown hooks: send SIGTERM under load and confirm readiness flips before active request count starts falling, then confirm in-flight handlers finish within the deadline while new connections are refused. Watch restart error rate and incomplete response count for one roll — a spike usually means a long handler ignored the abort signal or a pool close raced the last write. Keep a canary with a longer drain beside the new deadline for one deploy so you can tell a tight timeout from a real dependency hang.