...
- The task (StreamTask, StandbyTask) registers its state stores. State stores load offset metadata from the checkpoint file (link). That step aims to establish a mapping between data in the state store and the offset of the changelog topic.
- In case of crash failure, if the state store has data, but the checkpoint file does not exist. , ProcessorStateManager throws an exception in that case for EOS tasks. This is an indicator to throw away local data and replay the changelog topic (link).
- The task processes data and writes its state locally.
- The task commits EOS transaction. TaskExecutor#commitOffsetsOrTransaction calls StreamsProducer#commitTransaction that sends new offsets and commits the transaction.
- The task runs a postCommit method (StreamTask, StandbyTask) that:
- The task shuts down. It stops processing data, then writes its current offset to the checkpoint file and halts.
...