Here is the log from #18235; if you click Show Previous Attempts on the side bar, you can find logs from the crashing runs. Apparently someone else also restarted it 3 more times, too.

I don't have access to macOS to test so I might not be much help. Re-running the full test suite on my local machine, I did notice that the test leaves two windows hanging around after. This doesn't happen on a partial run of test_backends_interactive.py, so there might be some spooky interaction at a distance. Perhaps we could try running the test in a subprocess?

Thanks for looking at this @richardsheridan !

Learned two things from these PRs:

1. Subprocess isolation works to keep the test worker from crashing at least. In the stderr of those tests on macos it seems it can't find configure_subplots even though it's clearly part of the declared interface.
2. Taking a little care to destroy tkinter when you're done is enough to fix the weird long-lived windows when I run the full test suite locally, but it doesn't fix the crashes on macos. Unless we're prepared to go full cpython test suite on it, I'd say the subprocess strategy is the way to go.

I think we'll need someone with a mac and some experience with _macosx.m to debug missing attribute problem.

You should not be seeing NavigationToolbar2Mac if you are running with TkAgg. The subprocess is using the default backend, so it's not testing what you think it's testing. Set 'MPLBACKEND': 'TkAgg' in env to force it to use TkAgg.

On my mac, running 10.15.6 and Python 3.8, I can't get test_never_update to fail. I also can't get test_interactive_backend[toolbar2-macosx] ever to pass...