Do I have to disable my current X display now? That seems like asking a bit much, but maybe I'm missing a configuration somewhere...

pip install pytest-xvfb
pytest lib/matplotlib/tests/test_backend_qt.py

produces this error

INTERNALERROR>     raise XStartError(
INTERNALERROR> pyvirtualdisplay.abstractdisplay.XStartError: No success after 10 retries. Last stderr: b"_XSERVTransmkdir: ERROR: euid != 0,directory /tmp/.X11-unix will not be created.\n_XSERVTransSocketUNIXCreateListener: mkdir(/tmp/.X11-unix) failed, errno = 2\n_XSERVTransMakeAllCOTSServerListeners: failed to create listener for local\n(EE) \nFatal server error:\n(EE) Cannot establish any listening sockets - Make sure an X server isn't already running(EE) \n"

then I uninstall and the test suite runs again with no error.

Do I have to disable my current X display now?

I've only used it out of the box and as I understand, the whole point is that it will detect if xvfb is available and if so start it, if required. If not, it should just go on as normal. No problems on CentOS (without xvfb installed) nor Arch (with xvfb installed).

It may be that this is an OSX issue? It seems like pyVirtualDisplay is tested on both 10.15 and 11, but installs xquartz. We are using it on the CI for OSX as well, without installing xquartz.

https://stackoverflow.com/questions/65890804/xstarterror-in-pyvirtualdisplay

(But not so interesting answer...)

Interesting, it may have been something wrong with an older version of xquartz that I had installed locally. I updated that and now I don't get any errors, but I also don't get the windows being hidden... They still pop up in front of my active window and take priority. This could be because I only have the command xvfb and not xvfb-run...
My take on it is that this is a "nice to have" feature for Linux, but not required and could cause issues for new contributors trying to run the tests (myself included!) if their Xwindows aren't set up properly. I think adding this package as a suggestion to the installation docs might make sense instead of adding it to the environment file? https://matplotlib.org/stable/devel/dependencies.html#optional-dependencies

might make sense instead of adding it to the environment file?

It is there: https://matplotlib.org/devdocs/devel/dependencies.html#dependencies-for-testing-matplotlib

The origin of this PR is that all those are in the environment except for pytest-xvfb, e.g. pytest-xdist. But, indeed, if it causes problems on some machines it is not a good idea to add it.

(I realize that xarray is also missing, but since I did the PR adding xarray as a test dependency I am not surprised I missed to add it to environment as well...)

Unfortunately, selectors are not supported in environment.yml conda/conda#8089

There is the pip hack: install pytest-xvfb using a pip section in case of Linux. Should I change to that or just close this PR?

I think pip (with a comment) is viable. If that's too messy, simply add the condo package but commented out, plus a hint that Linux/win users could activate it.

I'm confused where this PR is. @timhoffm has approved, but the last comment seems to be asking for a change?

I close this. It is maybe a bit inconsistent to not have it in environment, but no harm in skipping it and as it apparently breaks on some Mac installs that seems like the best way.