SEZ9 commented on PR #13060:
URL: https://github.com/apache/seatunnel/pull/13060#issuecomment-6905483611

   @zephyr45 Thanks for the detailed recheck against `7c272c5a` and for 
rerunning the regressions locally.
   
   - **F1/F2:** Removing `validateRestoredStateOwner` and the modulo rejection, 
and parsing each restored state with its own saved prefix, is the direction I 
was asking for. The contiguous-assignment, mixed-prefix and `3 → 4 → 3 → 4 → 3` 
scenarios you describe are the right coverage. I'll confirm against the diff in 
`7c272c5a` before closing these.
   - **F4:** Using the stable index (Spark partition ID) for the Kafka writer, 
with the broker retry test showing task IDs 99 and 100 sharing stable index 0 
and the abandoned record staying invisible under `read_committed`, addresses my 
concern. I'll verify in the diff and then close it.
   - **F7:** The option row plus semantics paragraph covering the ID shape, 
Zeta/Flink stateful recovery and the Spark speculation restriction (with ZH 
matching) is what I asked for. Same here: I'll check the docs change in the 
diff before marking it resolved.
   - **F3:** I accept that the fresh-writer path should not scan full history 
given the growing startup cost you observed across Spark micro-batches, and I 
agree that aligning to a checkpoint counter would not identify transactions 
that were never captured in checkpoint state. Your comment appears to have been 
cut off at "Flink's adapter counters ca…" — could you finish that thought? If 
the conclusion is that the `transaction.timeout.ms` window is the accepted 
bound for this case, please state that limitation explicitly in the Kafka sink 
docs next to the semantics text so users set the timeout deliberately. With 
that, I'm fine closing F3 as a documented limitation rather than a blocker.
   
   One item from the earlier review was not covered in your update:
   
   - **F5/F6:** the legacy shared-namespace fencing performed concurrently by 
every restoring writer, where sibling probes bump epochs and extend each 
other's `getEpoch()==0` scan. Could you confirm whether `7c272c5a` changes 
anything here (e.g. bounding the probe chain or having a single writer own the 
legacy fence), or whether you consider it acceptable as a one-time migration 
cost? Either answer is fine; I just need it stated on the thread.
   
   Once the F3 wording and the F5/F6 answer are in, and I've verified the above 
against the diff, I have nothing else outstanding on this PR.
   
   <!-- streview-comment:2377 -->


-- 
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]

Reply via email to