Rectangle mpl selector CI failure

[larrybradley]

The TestRectanglePixelRegion.test_mpl_selector_drag CI test started failing last night (after I merged #392), but it fails only on macos python 3.9 (py39-test-alldeps). I cannot reproduce this failure locally -- all tests pass when running this tox test on my mac.

There is a KeyError: 'draw_event' being raised from matplotlib, see
https://github.com/astropy/regions/runs/5054472167?check_suite_focus=true.

CC: @dhomeier

[dhomeier]

I can reproduce it with 3.5.1, only in my case it always fails on the 'key_release_event' signal instead of

        self._func_cid_map.setdefault(signal, {})
        proxy = _weak_or_strong_ref(func, self._remove_proxy)
>       if proxy in self._func_cid_map[signal]:
E       KeyError: 'draw_event'

Looks like the signal is removed right again after setdefault (erratically?) in self._remove_proxy; the test also only fails when I am running at least the full TestRectanglePixelRegion, not just the mpl_selector ones – and obviously the virtually identical TestEllipsePixelRegion passes without problems.
It's also passing again with the current HEAD 3.6.0.dev1530, but I cannot fathom what changes might have fixed it.

[larrybradley]

New data point: in the cron CI run last night, the macos py39-alldeps test passed, but the windows py39-alldeps test failed (https://github.com/astropy/regions/runs/5088806452?check_suite_focus=true). Another heisenbug?

[larrybradley]

Interestingly, the CI tests have all been passing since I merged #424, even though that seems completely unrelated.

[larrybradley]

Closing this now since the CI has been passing -- will reopen if it reappears.

[larrybradley]

The CI failure is back: https://github.com/astropy/regions/runs/5227025176?check_suite_focus=true#step:6:230

[larrybradley]

Haven't seen this error in almost 2 years, so closing.