I'm unclear whether removal is ok. The original purpose ofenvironment.yml is to provide contributors with an easy way to set up a dev environment. We should how complete that should be in terms of optional dependencies, but I'm inclined to make a dev environment as complete as possible.

yeah, that was a point I wanted some clarification on as well,wx was failing some tests and wasn't installed in the pip-main version of appveyor, so I removed it mostly to see if it solved all of the failing tests... It got most of them, but there is still one for GTK and the fontconfig one.

I suppose one alternative is to create a new CI environment.yml instead that strips these out. (or figuring out why the interactive tests are failing in the first place, but that is a bit more scope than I really want to take on in a "get the CI running again" PR, especially if its windows specific as I do not have a windows box at the ready for local testing currently)

the wx tests which are failing are all using thewx backend (rather thanwxagg orwxcairo) which is deprecated anyway (and has been since 2.0...)

I think we just skip those tests, keep the environment file in tact.

That leaves only the the gtk one which is failing.

This unskips the following test. Apparently fontconfig does not work.

__________________________ test_get_fontconfig_fonts __________________________[gw0] win32 -- Python 3.8.13 C:\Miniconda3-x64\envs\mpl-dev\python.exe    @pytest.mark.skipif(not has_fclist, reason='no fontconfig installed')    def test_get_fontconfig_fonts():>       assert len(_get_fontconfig_fonts()) > 1E       assert 0 > 1E        +  where 0 = len([])E        +    where [] = _get_fontconfig_fonts()lib\matplotlib\tests\test_font_manager.py:78: AssertionError

Eh, not exactly,fontconfig is actually a transitive dependency anyway, so that test was failing and I triedadding it explicitly to see if thatfixed it (it did not) because the error message indicated that it might be a version thing. Didn't get around to removing it again.

fontconfig is brought in in a couple ways with this environment.yml: (and additional ones on unix systems, which made it harder to track down)

but notably, graphviz transitively depends on fontconfig viapango and throughlibgd (graphviz directly depends on fontconfig on unix)

Some of the gtk/cairo may also depend on it transitively, though those were mostlyhost dependencies notrun dependencies from what I saw.

Okay, the crux is that_get_fontconfig_fonts relies on subprocess call tofc-config --help (to determine if a supported version is available)

but the windows version of this utility does not use posix-style options. (At least I think that's why, testing out windows style options now)

actually, looking at where_get_fontconfig_fonts is called, it doesn't get called on windows, so I just explicitly skipped that test

Yes, I don't think there's any point in running this test on Windows, asfontconfig is not used there.

Likely

because we don't want to mix conda forge and main channel.

Conda forge is actually configured by the environment, so everything is conda-forge. The earlier instance of installing win32 could actually remove the explicit-c conda-forge

https://ci.appveyor.com/project/matplotlib/matplotlib/builds/45337448/job/aie7c1u8m0pf3nf6#L1893

Where does conda say that the channel config from an environment.yml is preseved? I just tested with a minimal

name: cftestchannels:    - conda-forgedependencies:    - python=3.7

and when I do aconda install numpy in that environment,conda wants to install from the main channel.

In that case, its probably because we prepend conda-forge to the base channels list before creating the environment:

Ah, that should be it.

Is this and the following from a code formatter? My understanding is that we don't want code churn like this. If we want to introduce more strict formatting, we need to decide this centrally and introduce this in a coordinated way.

Guilty... I did applyblack (only to this file) when flake8 started complaining about the formatting of the module name filter

I can revert the quote changes if we want.