You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
test(tiers): a worker crash in the tier guard's xdist probe test interrupts the whole session #1756
Describe the bug
On #1755's GraalPy experimental leg (job 116215824320), the unit tier now runs to the end: 69 failed, 3298 passed, 153 skipped … in 5509.22s. Even so, pytest exits 2 (interrupted). The cause is the tier guard, not a real tier violation:
! xdist.dsession.Interrupted: 1 read(s) of a generated sample capture from a unit-tier test module; see tests/_tiers.py for the tier rule.
The read it reports is xdist_probe_generated.pcap. That capture is the deliberate probe of tests/test_tier_guard_xdist.py, not a read by real test code. The same run shows worker 'gw10' crashed while running 'tests/test_tier_guard_xdist.py::XdistSubprocessTests::test_xdist_run_reports_the_diagnostic_and_fails', one of 13 worker replacements after #1753's ExitOnOutOfMemoryError.
Expected behavior
A worker crashing in the middle of the guard's own probe test fails that test only. It should not leave the probe's read behind for the outer session's guard to report.
Additional context
Inference, not yet measured: the probe test is not crash-safe. Whatever it does to stop its subprocess's read reaching the outer session's guard does not happen when the worker dies mid-test.
It shows up on GraalPy only because that is the only leg whose workers crash. The CPython unit legs exit 0.
The job's check is still green (continue-on-error). The defect makes the leg report an exit 2 that is not about the tests.
Activity
JarryShaw commentedon Oct 15, 2026
A worker is on this.