Sink

Where a context's committed state events go: the Kafka Sink's topics, a warehouse's tables. The runtime's sink subscription reads the log, builds the StateEventRecords and hands them to write; it owns the semantics around the write (checkpoints, retries, parking), so a sink only writes.

• Durable on return: write returns once every record is durable where it goes (a broker's acknowledgement, a committed insert). The subscription moves its checkpoint only after that, so a crash never loses a record.

• Idempotent by id: publishing is at least once: a record may be written again after a crash, a retry or a replay, and an old owner of a partition may run beside the new one. Writing a record whose StateEventRecord.id was written before changes nothing a reader sees beyond a duplicate it drops by that id.

• Order: the records of one aggregate reach write in version order, one batch after the previous one has returned; a batch may hold records of several aggregates.

• Failure: write throws when not every record is durable; the subscription retries, then parks the aggregate. Some records of a failed batch may have been written already: the retry writes them again.