rmannibucau opened a new pull request, #9374:
URL: https://github.com/apache/hop/pull/9374
## Summary
Fix a monitor deadlock during local pipeline error shutdown, and prevent
reentrant/concurrent calls from stopping the same pipeline repeatedly.
A JVM thread dump showed this three-thread cycle:
```text
TableInput completion:
holds transform monitor
-> waits for pipeline monitor in the transform-finished listener
FilterRows completion:
holds pipeline monitor
-> waits for transforms-list monitor in transforms.size()
ExecSqlRow error shutdown:
holds transforms-list monitor in stopAll()/synchronizedList.forEach()
-> waits for TableInput monitor in stopRunning()
```
The SQL exception in `ExecSqlRow` triggers shutdown; the deadlock is in the
shutdown/completion locking, not the SQL execution itself.
## Changes
- Iterate snapshots of the transforms in `stopAll()`, `safeStop()` and
`killAllNoWait()`, so transform callbacks do not run under the synchronized
transform-list monitor.
- Atomically claim `stopAll()` before invoking callbacks, preventing
duplicate concurrent shutdown and recursive shutdown through transform
callbacks.
- Keep terminal-state updates in `BaseTransform.markStop()` synchronized,
but notify finished listeners after releasing the transform monitor.
- Snapshot the finished listeners and invoke callbacks without holding their
list monitor. Listeners added during notification participate only in
subsequent notifications, rather than invalidating the current iteration.
- Move error-triggered stopping out of the pipeline monitor in the
completion listener.
## Regression coverage
Seven new bounded regression tests cover:
- Access to the transform list from another thread while `stopAll()` and
`safeStop()` execute transform callbacks.
- Reentrant `stopAll()` from a transform stop callback.
- Concurrent `stopAll()` calls notifying stopped listeners once.
- Safe stopping only input transforms, allowing downstream processing to
drain.
- Completion callbacks acquiring the transform and listener-list monitors
from another thread.
- Concurrent/reentrant `markStop()` and listener registration during
completion.
## Validation
Verified with Java 21 on the original development checkout before applying
the identical four-file patch to this branch:
- `PipelineTest`: 23 tests, zero failures/errors.
- `BaseTransformTest`: 19 tests, zero failures/errors.
- The broader engine test/package invocation exceeded the 125-second command
timeout. Its completed Surefire reports contained 1,054 tests across 176
reports with zero failures/errors; this is **not** a successful full-suite
result. The log also included a background timer exception about an
uninitialized central log store; its relationship to the timeout was not
established.
- Subsequent engine packaging with `mvn -q package -DskipTests
-Dassembly.skipAssembly=true` succeeded.
- `git diff --check` passed; the patch applied cleanly to current `main`.
The initial test-compilation attempt found a missing import in the new
tests, which was corrected before the above results. Tests have not been rerun
on this rebased branch. The production workload and its failure-shutdown path
have not been replayed with the patched runtime.
This branch contains only the deadlock fix and tests, not the separate Table
Input JDBC partitioning work.
---
### Contribution checklist
- [ ] Run `mvn clean install apache-rat:check` (not performed; validation
limits above).
- [x] Keep related changes in a single commit.
- [ ] Mention an issue if applicable (no separate issue filed).
- [ ] I hereby declare this contribution to be licensed under the [Apache
License Version 2.0, January 2004](https://www.apache.org/licenses/LICENSE-2.0)
(for contributor confirmation during review).
--
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]