pytest-xdist "worker ... crashed" / "Replacing crashed worker" in CI

With pytest -n (xdist), tests run in worker subprocesses. When a worker dies hard - a native segfault, an OOM kill, or a test calling os._exit - xdist reports the worker crashed and may replace it, often losing the test that caused it.

What this error means

The run prints [gw0] node down: Not properly terminated and replacing crashed worker gw0, or worker gw1 crashed while running. The failure can be flaky and hard to attribute because the offending test’s output is lost with the worker.

pytest output

[gw0] node down: Not properly terminated
replacing crashed worker gw0
INTERNALERROR> ... worker 'gw0' crashed while running
'tests/test_native.py::test_segfault'

Common causes

A native crash (segfault) in a C extension

A C/Rust extension or ctypes call segfaults the worker process. Python cannot catch it, so xdist sees the worker die.

Out-of-memory kill on a parallel worker

Running too many workers multiplies memory use; the OOM killer terminates one, and xdist reports it as crashed.

A test calling os._exit / hard-killing the process

Code under test (or a misbehaving fixture) calls os._exit or sends a fatal signal, terminating the worker without normal teardown.

How to fix it

Isolate the crash by running serially

Re-run without xdist to get a clean traceback and identify the test.

Terminal

pytest -p no:xdist tests/test_native.py -x
# or reduce parallelism
pytest -n 2

Address the root cause

Segfault → pin/upgrade the native dependency, or run that test in isolation.
OOM → lower -n (workers), or raise the runner memory; watch total = per-worker × workers.
os._exit → fix the code/fixture so it raises or returns instead of exiting.

How to prevent it

Size -n to the runner’s memory, not just its CPU count.
Keep native dependencies pinned and tested for stability.
Quarantine known-crashy native tests rather than letting them flake the suite.