Write down who owns the retry before go-live
Night-time failures get expensive when both sides assume the other will re-send. Before go-live on an internal API hop, write a single sentence: who owns the retry, how many attempts, and where a dead letter lands. Put that sentence in the handover note where on-call staff will actually find it.
Ambiguity usually looks polite in workshops. It looks costly at 02:00 when finance is missing receipts and warehouse staff have gone home. Health checks we run after release seasons often find retries implemented on both ends—or on neither—because nobody wanted to look uncooperative in the planning meeting.
Assigning ownership is not about blame. It is about making the path operable. If the producer retries, the consumer must be idempotent. If the consumer polls, the producer must keep a recoverable outbox. Pick one model and document it before the cutover window opens.