As I wrote on #10661, I wonder if we shouldn't replace this logic entirely to just use os.path.expanduser('~') as the default on both POSIX and Windows systems. By doing so we can delegate this logic to what Python already provides, which is probably better battle-tested and thought out by experts. According to the docs:

On Windows, HOME and USERPROFILE will be used if set, otherwise a combination of HOMEPATH and HOMEDRIVE will be used. An initial ~user is handled by stripping the last directory component from the created user path derived above.

It appears HOMESHARE is not used at all in this logic. According to this page which documents Windows environment variables better than anything I could find directly from Microsoft, HOMESHARE is only meaningful if HOMEDRIVE happens to be a UNC path (and I guess in this case they would be the same), in which case HOMESHARE points to directory under HOMEDRIVE that contains the user's home. A lot of this probably exists for legacy purposes that aren't very meaningful to a typical Windows user outside some specialized enterprise deployment.

Also worth noting that as of Python 3.8:

Changed in version 3.8: No longer uses HOME on Windows.

That change was motivated by this issue.

@embray , I tried to implement your suggestion in the second commit but there is a chance I completely misunderstood, so please review. Does the doc at https://docs.astropy.org/en/latest/config/index.html need updating too?

@pllim Sorry for the late reply. The last week has been kind of crazy for me; I will be around more after this week.

Looking at the documentation, it looks like it conflicts with my previous recommendation that Windows users not use XDG_CONFIG_HOME, but now I can see where someone might have gotten the idea to if their "default" home directory was not being found properly.

I would argue we change that recommendation, but actually in retrospect there is not currently any other way to specify a different astropy config path via the environment except for setting XDG_CONFIG_HOME, since not every system uses the XDG base directory specification--depending on this is a bit system-biased (of course it's just a specification that one can in principle use on any system).

Perhaps we should take a cue here from matplotlib's MPLCONFIGDIR. I will open a separate issue for this, if there isn't already.

I'm otherwise 👍 on this as-is.

@embray , thanks for opening the meta issue, it was really illuminating. If you are okay with this, would you mind approving and merging? If you need write access, please let Kelle know.

Cool, looks like I have access to merge. Just had to approve my review.

Thanks! 😄