SEZ9 commented on PR #13060: URL: https://github.com/apache/seatunnel/pull/13060#issuecomment-6936149729
@zephyr45 thanks, that answers my F3 question. **F3:** Understood. Flink's adapter counters can differ between writers, so they can't reliably identify an abandoned transaction that never reached checkpoint state. A fresh writer fences sequence 1, but a later abandoned transaction can still block `read_committed` until `transaction.timeout.ms` expires. Since you point to both the EN and ZH guides at `7c272c5a` stating this next to the recovery semantics, I'm fine treating it as a documented limitation for this PR. **F5/F6:** Also understood that the behavior is unchanged. Deduplication is per writer only, the shared legacy prefix has no single owner, and the scan has no new bound. Writers can therefore still bump each other's epochs and extend the scans, and a retried migration can repeat that work. I can accept this as a migration cost here with coordinated cleanup deferred. Please open the follow-up and link it in this thread so it stays tracked. You asked whether you missed anything. Your comment covers F3 and F5/F6, but I can't tell from it where the other points stand at `7c272c5a`. Could you confirm each briefly: - **F1:** Does a Flink scale-down restore of an EXACTLY_ONCE Kafka sink now recover, given that Flink redistributes ListState in contiguous chunks rather than by modulo? - **F2:** Does a scale-down restore work when `transaction_prefix` is not configured, where each writer generates its own random prefix? - **F4:** On Spark, is the previous attempt's open transaction fenced after a task retry, or is this also a known limitation? If it is a limitation, where is it documented? - **F7:** Is the `transaction_prefix` option row aligned with the new ID-shape description, and is the engine scope of the rescaling claim stated in the docs? A one-line pointer to the relevant change or test for each is enough. <!-- streview-comment:2390 --> -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
