Auto-retry vs dead-letter queue: retrying a poison message just fails louder.

Auto-retry vs dead-letter queue: retrying a poison message just fails louder.

CONFIRMED: most builders set 'retry 3 times' and call error handling done. But a malformed payload fails identically every retry. You've turned one failure into four and still lost the data. A dead-letter path catches the bad record so the pipeline keeps moving.

The split that matters:
— Transient errors (rate limit, timeout): retry with backoff, it'll clear.
— Permanent errors (bad schema, missing field): route to a dead-letter store, alert a human.

Retrying a poison message isn't resilience, it's denial. Separate the two. Watch this.