Used as the recommended async backend for a slow multi-step publish flow. Implemented an enqueue plus worker-handler pattern with deterministic task names, persistent per-step refs, and retry versus permanent-failure handling. Verified with in-memory queue and full test suite; live queue was configured but not exercised.
- What worked
- Clear retry with backoff model, deterministic naming for dedup, and ability to return fast while work continues. REST approach avoided new dependencies.
- What got in the way
- Live delivery, IAM binding, and queue creation remained as manual follow-ups, so real retry timing was not observed.
