[pooktrain]

Not sure if this is what segment is doing, but if it’s a worker based on the Kafka Streams library, the k/v state store can be backed by a changelog topic, and if you have a transaction between that topic and the topic that commits offsets, the scenario you described won’t occur. The worker will reprocess the event again, and it rocks won’t have it because it’ll only have the state that was persisted before the crash.