Atomic but Not Magic

A Kafka read-process-write job uses a transactional producer, writes results to two output topics, commits its input offsets inside the transaction, and downstream consumers read with isolation.level=read_committed. What does this actually guarantee?

The job's own handler code is guaranteed to execute exactly one time per input record, so no processing step can ever run twice after a crash

The two output writes and the offset commit become visible all together or not at all, giving exactly-once processing for the Kafka-to-Kafka path

Any external HTTP call or foreign database write the job makes during processing is enrolled in the same transaction and therefore also executed exactly once

Records inside one transaction are delivered to consumers in a single global order spanning both output topics, so cross-topic sequence is preserved