A dead-letter queue is filling up. How would you investigate and recover safely?

In production DLQ tooling supports filtered preview, payload redaction, reason grouping, small-batch replay, archive, and an audit trail of every operator action. Replayed messages keep their original identity and correlation metadata so idempotency still works and investigation remains connected. Recovery begins only after the handler, data, or contract defect is understood, then proves a sample before increasing batch size. Alerts use dead-letter rate and oldest age, and the incident closes with a regression test or producer validation change for that failure class.