SEZ9 commented on PR #12134:
URL: https://github.com/apache/seatunnel/pull/12134#issuecomment-6871184035

   Thanks for the follow-up on head `c6fd7fc0d575c9c316fb41beb94eb1fd6dcaefa0`. 
The description update is helpful: documenting that the redrive deadline is 
only checked between attempts (not bounding a blocking `savePoint().get()`), 
and that the heap-only intent can be lost in a second failover before 
`DOING_SAVEPOINT` is persisted again, makes the single-failover scope explicit. 
I also see that fork Build 41162516391 attempt 2 succeeded and that Apache now 
reports Build success on the same `c6fd7fc0` head.
   
   That said, a green build and an accurate description don't by themselves 
close the review points on `CheckpointCoordinatorFailoverIT.java`, and the head 
is unchanged, so the following remain open:
   
   1. **Production-side changes.** The `CoordinatorService.java` / 
`JobMaster.java` changes that the test's terminal 
`SAVEPOINT_DONE`-after-failover assertion depends on still aren't visible in 
what I can review. Please make sure the full diff is reviewable, or summarize 
in the PR where the redrive is triggered and how the pending savepoint intent 
is picked up by the new master.
   
   2. **Race between probe and master kill.** Detecting the in-flight savepoint 
and then calling `masterNode1.shutdown()` leaves a window where the savepoint 
finishes on the old master, so the test can pass without exercising the redrive 
path. Please add a guard that fails the test if the savepoint completes before 
the shutdown is observed.
   
   3. **Reflection polling robustness.** Polling `CheckpointCoordinator`'s 
private `pendingCheckpoints` field via reflection can throw mid-transition and 
abort the Awaitility wait. Please add `ignoreExceptions()` or equivalent 
defensive handling. Longer-term, since this PR is modifying coordinator 
internals, consider a package-visible accessor or test hook so the regression 
guard isn't coupled to a private field name.
   
   4. **Executor and future lifecycle.** `savepointCaller` is created outside 
the try block and its shutdown isn't visible, and the `savepointCall` future is 
never inspected. A blocked `savePointJob` thread can leak, and a pre-failover 
savepoint failure would surface only as an unrelated 30s Awaitility timeout. 
Please shut the executor down on exit and wait on the future with a timeout so 
unexpected exceptions fail the test with the real cause.
   
   Once a new head addresses these, I'm happy to re-review promptly.
   
   <!-- streview-comment:2345 -->


-- 
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