The idea
A repeatedly failing message should not prevent unrelated work from progressing. A dead-letter queue separates persistent failures from the normal flow while retaining enough information to understand them. It is a work queue for recovery, not permanent storage.
Design choice
Store the failure category and original identity, and define who reviews the queue. Limit retries before diversion. Replay only after the underlying issue is understood, using the same duplicate-handling policy as the main flow.
Worked example
A message with an unsupported schema enters a review queue. After a compatible reader is deployed, an operator replays a small sample before releasing the remainder.
Before you ship
- Assign an owner and review interval
- Preserve failure context safely
- Test a controlled replay
Examples are illustrative. Adapt the checks to your own environment.