A read-process-write application can place output records and consumed offsets in the same Kafka transaction. A committed result then advances both sides together. Without that boundary, a crash can publish output while leaving the input position behind, causing replay after restart. This guarantee applies to Kafka records and offsets in the transaction; it does not automatically include an external database or an email. These tasks stage an output record and the input next offset before committing. Inspect both after the commit. The local transaction control exposes the result without implementing the complete client protocol. An answer that publishes output but forgets the offsets leaves the recovery gap open.
Inspect the topics, partition logs and group offsets before changing the cluster.