The "PR test summary (non-blocking)" job (ci/utils/pr_test_summary.py, wired up in .github/workflows/pr.yaml) only depends on and reports conda-cpp-tests, conda-python-tests, wheel-tests-cuopt, wheel-tests-cuopt-server, and test-self-hosted-server (_TEST_PREFIXES in the script, and the pr-test-summary job's needs: list). java-build isn't included, so it neither gates the summary nor shows up in it — the summary can post "all tests passed" while java-build is still running or has failed.
Possible fix
- Add
java-build to pr-test-summary's needs: list in .github/workflows/pr.yaml
- Add a matching prefix to
_TEST_PREFIXES in ci/utils/pr_test_summary.py
- Note:
_analyze_job_log only parses gtest ([ FAILED ]) and pytest (short test summary info) failure formats. A failed java-build job would still show up correctly as a failed job in the top-line counts, just without a per-test breakdown (JUnit/Surefire output isn't parsed) — that's an acceptable degrade, not a blocker, unless we also want per-test Java failures listed.
Context
Noticed while investigating #1820 / #1823: PR #1823's "PR test summary (non-blocking)" check passed while java-build / build was still in progress, which was confusing to reviewers since it looked like the whole PR was green.
The "PR test summary (non-blocking)" job (
ci/utils/pr_test_summary.py, wired up in.github/workflows/pr.yaml) only depends on and reportsconda-cpp-tests,conda-python-tests,wheel-tests-cuopt,wheel-tests-cuopt-server, andtest-self-hosted-server(_TEST_PREFIXESin the script, and thepr-test-summaryjob'sneeds:list).java-buildisn't included, so it neither gates the summary nor shows up in it — the summary can post "all tests passed" whilejava-buildis still running or has failed.Possible fix
java-buildtopr-test-summary'sneeds:list in.github/workflows/pr.yaml_TEST_PREFIXESinci/utils/pr_test_summary.py_analyze_job_logonly parses gtest ([ FAILED ]) and pytest (short test summary info) failure formats. A failedjava-buildjob would still show up correctly as a failed job in the top-line counts, just without a per-test breakdown (JUnit/Surefire output isn't parsed) — that's an acceptable degrade, not a blocker, unless we also want per-test Java failures listed.Context
Noticed while investigating #1820 / #1823: PR #1823's "PR test summary (non-blocking)" check passed while
java-build / buildwas still in progress, which was confusing to reviewers since it looked like the whole PR was green.