Replay is the concrete capability behind design for replay. Because every canonical message is persisted and tracked when published, the bus holds a history you can draw from. Fix the bad map, register the missing subscription, or bring the endpoint back, then select the affected messages, from the archive or the dead-letter store, and re-publish them.

Replay is only safe if reprocessing is safe, which is why idempotency is a standing rule: a replayed message must not double-apply. Done right, recovering from an outage or a bad deployment is routine (select the window, replay, done) and it is the same machinery that lets event sourcing rebuild state from history.