After a bounded number of retries (usually with backoff), a poison message is republished to a DLQ stream together with its error context. Engineers can inspect and replay it later without losing data or stalling healthy messages.