Seen on the first Windows legs to get past setup after gh-ocannl-935 (dispatch runs https://github.com/lukstafi/ocannl-staging/actions/runs/34262512499 and 34266955741). The benchmarks/dune rule that runs python3 test_orchestrate.py fails one case on windows-latest:
FAIL: test_a_child_killed_mid_stream_preserves_its_partial_stdout (test.test_cell_group.CellGroupTest.test_a_child_killed_mid_stream_preserves_its_partial_stdout)
File "…\_build\default\benchmarks\test\test_cell_group.py", line 113
self.assertEqual(result.stdout, b"partial child output")
AssertionError: '' != b'partial child output'
Ran 209 tests in 19.763s — FAILED (failures=1, skipped=17)
(the second run additionally reported errors=2, so the case is not even deterministic there). The orchestrator's cell group kills a child mid-stream and expects the bytes it had already written to survive; on Windows the kill path (TerminateProcess through Popen.kill, or the process-group signal the POSIX path uses) apparently discards the pipe contents, or the child is killed before its write is flushed. 17 cases are already skipped on Windows, so the suite has a platform-skip mechanism; the question is whether this case's contract (partial stdout survives a kill) is achievable on Windows or should be skipped there with the reason recorded. Filed by the wave coordinator from the gh-935 verification runs.
Seen on the first Windows legs to get past setup after gh-ocannl-935 (dispatch runs https://github.com/lukstafi/ocannl-staging/actions/runs/34262512499 and 34266955741). The
benchmarks/dunerule that runspython3 test_orchestrate.pyfails one case onwindows-latest:(the second run additionally reported
errors=2, so the case is not even deterministic there). The orchestrator's cell group kills a child mid-stream and expects the bytes it had already written to survive; on Windows the kill path (TerminateProcessthroughPopen.kill, or the process-group signal the POSIX path uses) apparently discards the pipe contents, or the child is killed before its write is flushed. 17 cases are already skipped on Windows, so the suite has a platform-skip mechanism; the question is whether this case's contract (partial stdout survives a kill) is achievable on Windows or should be skipped there with the reason recorded. Filed by the wave coordinator from the gh-935 verification runs.