shuke987 opened a new pull request, #68981:
URL: https://github.com/apache/doris/pull/68981
### What problem does this PR solve?
Issue Number: N/A
Problem Summary:
When a PR receives another commit, its existing AI review can keep consuming
model capacity until Codex notices the stale head. It then fails final
submission verification with `No final review submission was declared`. This
also happens while the outer runner waits five minutes to resume after a
capacity error.
Poll the authoritative PR head every 60 seconds during execution and
capacity backoff, and check at attempt boundaries. A confirmed different SHA
stops the old attempt through the existing
SIGINT/grace-period/descendant-reaping path. API errors, timeouts, and
malformed responses only warn; they do not establish supersession. Existing
strict resume and submission checks remain in place.
Record `review.superseded`, both SHAs, and a final diagnostic. The workflow
validates the distinct helper exit code/result, treats supersession as a
handled outcome, skips verdict/status publication and obsolete status
aggregation, and retains Litefuse, session, and auth finalization. A successful
workflow in this case does **not** mean a passed review. Litefuse verification
also permits a review stopped before its first model/tool request. New commits
still require the existing `/review` trigger; this does not add automatic
reviews on pushes.
### Release note
None
### Check List (For Author)
- Test
- [x] Unit Test
- [x] Manual test (local process integration and workflow shell checks
below)
- Behavior changed:
- [x] Yes. Superseded review processes stop promptly and have an explicit
handled outcome.
- Does this need documentation?
- [x] No. Internal CI behavior; triggering a review is unchanged.
Validation:
- `python3 -m unittest discover -s .github/scripts -p 'test_*.py'` in the
isolated snapshot containing the runner, submission, and Litefuse test modules:
**138 cases**, all covered across macOS Python 3.12 and native Linux Python
3.13. macOS passes 135 and skips 3 Linux-only cases; Linux passes 137 and skips
the Bash/jq case, which passes on macOS.
- Linux runs with network disabled and a read-only source mount. Real
processes cover graceful and forced interruption, detached descendants,
repeated cancellation during cleanup, preservation of unrelated processes,
capacity resume, and head-change termination. GitHub and Codex endpoints are
local fakes; no production review was triggered.
- New negative cases cover transient API failures, malformed heads,
exhausted budgets, stale heads before start/at exit/during
verification/backoff, cleanup failure propagation, and mismatched workflow
result metadata. The Bash test executes the actual workflow result-handling
block with jq.
- `actionlint -shellcheck= .github/workflows/code-review-runner.yml` and
`git diff --check`: pass.
- A live new-commit scenario against the deployed workflow remains to be
validated after rollout. No production binary or credentials change is needed.
<details>
<summary>Self-review checkpoints</summary>
| Checkpoint | Conclusion |
| --- | --- |
| Goal | Stops work on a confirmed obsolete head; unit and Linux process
tests cover execution and backoff. |
| Scope | Six CI files; no Doris product source, runtime version, model, or
trigger changes. |
| Concurrency | Polling runs synchronously on the runner's main thread; the
existing stderr copier does not share polling state. No new locks or polling
thread. |
| Lifecycle | gh finishes before child reaping; exceptions unwind through
existing SIGINT, bounded wait, kill, reaper, and copier cleanup. Cleanup errors
remain errors. |
| Configuration | Fixed 60-second interval and 10-second API timeout,
bounded by the existing shared budget; no new user setting. |
| Compatibility | Existing CLI remains compatible. Helper and workflow
consume the same pinned workflow revision. The added result state has an
explicit workflow branch. |
| Thrift/protobuf | Not applicable. |
| Parallel paths | Initial attempts, model fallback, same-session resumes,
capacity waits, and verification races were checked. |
| Conditions | Only a valid, different head SHA signals supersession.
Base/open-state and duplicate-submission checks remain strict before resume. |
| Coverage | Unit, workflow shell, telemetry metadata, and real Linux
descendant tests cover normal and negative paths. Live production validation is
still outstanding. |
| Test results | Both final suites pass with complementary platform skips;
workflow syntax and whitespace checks pass. |
| Observability | Explicit result, final message, JSONL event, workflow
summary, and Litefuse metadata preserve the old/new SHA distinction. |
| Transactions/persistence | No product transaction changes; result writing
reuses the existing atomic JSON writer. |
| Data writes | Superseded runs skip source status and aggregation; existing
session/auth finalizers are retained. No new review/comment write is
introduced. |
| FE/BE propagation | Not applicable. |
| Performance | One bounded head lookup per minute during long
execution/backoff, plus boundary checks; avoids continuing obsolete model work.
|
| Remaining limits | Detection depends on a successful API lookup and is
periodic. It cannot retract a GitHub write already completed before the head
change was observed. |
</details>
### Check List (For Reviewer who merge this PR)
- [ ] Confirm the release note
- [ ] Confirm test cases
- [ ] Confirm document
- [ ] Add branch pick label
--
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]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
