SEZ9 commented on issue #12344:
URL: https://github.com/apache/seatunnel/issues/12344#issuecomment-5723587076
Quantifying the blast radius, since my previous comment argued it from the
workflow gate rather than from data. I surveyed the `Build` status of the 40
most recently updated open pull requests on 2026-09-17:
```
FAILURE 15
SUCCESS 12
IN_PROGRESS 6
ACTION_REQUIRED 3
CANCELLED 3
QUEUED 1
```
The split between the 15 red and the 12 green is not random — it lines up
exactly with the `if: api == 'true' || engine == 'true'` gate:
- **All 12 green PRs touch only connectors, docs, or connector e2e**
(#12312, #12315, #12323, #12324, #12325, #12326, #12328, #12329, #12341,
#12347, #12349, #12370). None of them sets `engine` or `api`, so
`all-connectors-it-2` is skipped and they never reach `testAddFieldWithRestore`.
- **Every Zeta/engine PR in the window is red**: #12311, #12313, #12314,
#12316, #12338.
I drilled into two of them at job level, and the failure lists match my own
PRs exactly:
```
#12316 FAIL all-connectors-it-2 (8, ubuntu-latest)
FAIL all-connectors-it-2 (11, ubuntu-latest)
FAIL all-connectors-it-7 (11, ubuntu-latest)
#12338 FAIL all-connectors-it-2 (8, ubuntu-latest)
FAIL all-connectors-it-2 (11, ubuntu-latest)
FAIL all-connectors-it-1 (8, ubuntu-latest)
FAIL all-connectors-it-1 (11, ubuntu-latest)
FAIL engine-v2-it (8, ubuntu-latest)
```
To be precise about what I checked: I verified per-job failures for 2 of the
15, not all 15. For the rest I have only `Build=FAILURE`.
The part that seems worth a maintainer's attention is the compounding
effect. Two of the blocked PRs are themselves fixes for the *other* flaky tests
currently failing alongside this one:
- **#12311** — `[Fix][Zeta] Resolve a CANCELING vertex to CANCELED, not
FAILED`, the fix for `SplitClusterFaultToleranceIT` (#12353). `Build=FAILURE`.
- **#12313** — `[Test][E2E] Make BackpressureSlowSinkIT barrier injection
deterministic`. `Build=FAILURE`.
So the flaky-test cleanup that would make `engine-v2-it` trustworthy again
cannot merge, because an unrelated Opengauss CDC test in a different job blocks
it. That is a fairly tight deadlock, and this issue is the only edge in it that
a single change can cut.
My earlier offer stands and I still don't want to act on it unasked: if
quarantining is preferable to a fix in the short term, I'll send the mechanical
`@Disabled` PR with a link back here. Equally happy to leave it if someone is
already looking at the restore path — I only want to make sure the cost of
waiting is visible.
--
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]
