Vivek1106-04 commented on PR #12165:
URL: https://github.com/apache/seatunnel/pull/12165#issuecomment-5807840859
@SEZ9 @DanielLeens on the three asks from #12339:
**1. Wording.** The class Javadoc, the STIP Correction and the STIP Test
Plan already carried it at `c773cc091`. The PR description did not, and it had
two claims the Correction made wrong. It is updated now:
- Section 3 no longer says a blocked body delays "nobody else's" work. It
says a dispatch thread is shared, and that what a dispatched body may block on
therefore matters.
- Section 4 now says the savepoint "wait until it drains" step is a
sleep-poll inside the pipeline's lock, and that this is unchanged `dev`
behaviour.
- A new section, "What this PR does not change: savepoint lock holding
(#12441)", states that this PR does not remove the coupling. It says the
coupling is bounded by the length of a savepoint, and that the regression lands
with #12441, written red first.
- A new section describes the watchdog expiry change from `da0cbbf52`, and
the testing section names
`testCheckpointExpiryHandlingNeverRunsOnTheDispatchThread`.
**2. Failed Build on `c773cc091`.** There were three failed jobs, with two
distinct failures. Both are open flakes on `dev`, and neither involves this
PR's code:
- `engine-v2-it (8)`:
`BackpressureSlowSinkIT.testCheckpointsKeepCompletingUnderSustainedBackpressure:247`,
"only observed 1 (samples=[1, 1, 2, 2, 2, 2, 2, 2, 2])". The same assertion
fails on `dev` itself (runs `34821874806` and `34801745423`, cited in #12313).
#12313 fixes the fixture and #12316 fixes the cause: the source reader
re-enters the checkpoint lock back to back and starves barrier injection.
Neither is merged. The job's own log shows the scheduler doing its part:
| time | event |
|---|---|
| 14:18:06.26 | periodic trigger armed, interval 15 s |
| 14:18:21.26 | checkpoint 1 triggered, 15.00 s later |
| 14:18:41.51 | checkpoint 1 completes, after 20 s |
| 14:18:41.77 | checkpoint 2 triggered (due at 36 s, found 1 pending, took
the 500 ms re-arm) |
| 14:18:58.61 | checkpoint 2 completes, after 17 s |
| 14:18:58.77 | checkpoint 3 triggered, still not complete when the 90 s
window ended |
Every trigger fired on time. What stalled was the barrier reaching the
source, which is the shape #12316 describes (20 / 34 / 8.5 / 36 s checkpoint
durations on `dev`).
- `unit-test (8, windows)` and `unit-test (11, windows)`:
`PayPalClientTest.closeWakesRetryWait:416` in `connector-http-paypal`. This PR
does not touch that module. #12370 already widened the test's budget, and
#12444 is open to fix the latch ordering it still races on.
I have re-run the failed jobs and will post the result here.
**3. #12441 handoff.** It is recorded there
([comment](https://github.com/apache/seatunnel/issues/12441#issuecomment-5806797054)),
with the red-first requirement. Part 2, the cross-pipeline regression, needs
the shared pool from this PR. So if #12441 is ready before this PR merges, it
lands with the same-coordinator part, and I will add the cross-pipeline
regression once this PR is in.
--
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]
