SEZ9 commented on PR #12134:
URL: https://github.com/apache/seatunnel/pull/12134#issuecomment-6888767937

   Thanks for the follow-up on `c91274d45993d5e14648bc4e4d75fcd64d4bfdf2`. 
Capturing the old `PendingCheckpoint` future and rejecting a successful 
acknowledgment both after shutdown and again after the new master persists its 
savepoint is a good step on the "did the test actually exercise failover" 
concern: the test now has a way to fail when the old master completes the 
savepoint instead of silently passing. Restoring the process deploy mode is 
also appreciated.
   
   A few items from the earlier review are still open on this head:
   
   1. **Production-side fix visibility.** I can still only review the E2E test 
in the diff body; the `CoordinatorService` / `JobMaster` changes that the 
terminal `SAVEPOINT_DONE` assertion depends on aren't visible to me. Please 
make sure those changes are included in the PR diff (or point me to where they 
are) so the redrive path can be reviewed alongside the test.
   2. **Detect-then-kill race.** The new post-shutdown rejection tells us when 
the test degraded into the "savepoint already finished" scenario, but it turns 
that into a failure rather than preventing it. Could you describe (or add to 
the test) what happens if the savepoint is not observed in flight before the 
master is stopped — e.g. a bounded retry of the trigger/kill sequence, or at 
least a clear assertion message so a flaky run is distinguishable from a real 
regression?
   3. **Reflection polling robustness.** You note the existing reflection 
polling is unchanged. Since it reads live coordinator internals that can be 
mid-transition, please add `ignoreExceptions()` (or equivalent defensive 
handling) to the Awaitility wait so a transient exception doesn't abort the 
test. A short comment acknowledging the coupling to `pendingCheckpoints` would 
also help future maintainers.
   4. **`savepointCaller` executor and `savepointCall` future.** I don't see 
these addressed in the update. Please make sure the executor is shut down on 
every exit path and inspect the future's outcome with a bounded wait (asserting 
on the expected result or exception), so a pre-failover savepoint failure 
surfaces directly instead of as an unrelated Awaitility timeout, and a blocked 
`savePointJob` thread can't leak.
   5. **CI on the new head.** The passing fork run 41162516391 and the Build 
success you mention are on `c6fd7fc0`; as you say, the applicable CI still 
needs to run on `c91274d45993d5e14648bc4e4d75fcd64d4bfdf2`. Once it does, 
please drop a note here with the result.
   
   Once the production changes are reviewable and items 3–4 are in, I'll take 
another full pass.
   
   <!-- streview-comment:2361 -->


-- 
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