An application often needs to update its database and publish a business event. A plain sequence of those two actions has a crash gap: one can succeed while the other does not. A transactional outbox writes the business change and an outgoing event within one database transaction, then a publisher or change-data-capture process relays that event to Kafka. The relay can repeat an event, so downstream consumers still need a stable event identity and a replay policy. These decision tasks distinguish that boundary from producer retry idempotence. No browser command can make two unrelated external systems share a transaction. Choose the design that states how the database change and repeated delivery are handled.
Inspect the topics, partition logs and group offsets before changing the cluster.