indeterminate live_server segfaults on macos (workaround: threading.Thread) #103
Labels
stale
This issue has not seen activity for a while
Comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Not sure this is short-run actionable, and I don't think it's actually a pytest-flask bug. If others encounter this it may still make sense to add something like the workaround I identified as an optional run mode or automatic fallback. Opening an issue in case it helps others or attracts more information/reports. The lack of similar reports makes me suspect this might come down to the OS or specific dependency versions--but I haven't had time to fiddle with them.
We had some trouble this fall with flaky test runs using live_server and selenium for end-to-end tests. I'll describe everything I can recall in case it helps someone find the issue; scroll to the end for the workaround. :)
Initial encounter/debug
The issue initially manifested as inconsistent test/fixture failures at the first location in a test that depended on finding/interacting with anything specific loaded (because the server never started and the browser has a blank page loaded). When I first encountered this, the test exception/assertion errors were all I got.
I eventually figured out that
live_server._processcould silently fail to start during setup. You can detect this with something like:While most of my test runs fail (running 2 parameterized tests 8x each), there's usually only one failure across the run. It often (but not always) fails on the same test. Any remaining tests usually run fine. Playing around for a bit (different numbers of tests, turning each test into a simple pass, renaming them to force different orders) made me think that where it breaks is somewhat correlated with how much work the tests make the server do.
Stepping through the code isolated the problem to when
app.run(host=host, port=port, use_reloader=False, threaded=True)is called inside the worker/target function passed tomultiprocessing.Process. At the time, I had trouble introspecting the child process (see the paragraph), but I did notice the crash produces OS crash reports. Here's an example:segfault crash report
Today I disabled my workaround to ensure I still see the issue and noticed that something (maybe pytest--I was using 4.x at the time and now use 5.x) in my environment has shifted and the test run now prints helpful Python segfault messages and full stack traces. Here's an example:
segfault stack trace
Workaround
At least in our case, threads proved a viable workaround. I achieved this by subclassing LiveServer and overriding the live_server fixture in our conftest:
The text was updated successfully, but these errors were encountered: