I just had a look at the sources ; in lib/matplotlib/init.py:

def _get_home():
    """Find user's home directory if possible.
    Otherwise, returns None.

    :see:  http://mail.python.org/pipermail/python-list/2005-February/325395.html
    """
    try:
        path = os.path.expanduser("~")
    except ImportError:
        # This happens on Google App Engine (pwd module is not present).
        pass
    else:
        if os.path.isdir(path):
            return path
    for evar in ('HOME', 'USERPROFILE', 'TMP'):
        path = os.environ.get(evar)
        if path is not None and os.path.isdir(path):
            return path
    return None

so it is looking into environment variables...

I agree that crashing strangely in os.path.join is bad, but what should it do if the home directory doesn't exist? Create it? I'm not sure that's the best idea either. If $HOME points to a non-existent directory, generally that means something is wrong (and we should probably raise a better exception, but I'm not sure blindly returning it is the answer).

Ooohh... I had misread the original bug report... I had understood the code wasn't finding the home even though it was explicit... and now I see that $HOME points to something invalid!

Well I have filed a report instead of a PR because for me all the ambiguity where TMP could be the "home" is not clear. Imho $HOME is the home if it exists or not. On POSIX systems I guess HOME variable is always there... If this all was due to platform differences -- then code should reflect it. Now it just looks like a heuristic to me to get some directory to write to... then may be it should mktemp directory and return it as home if none among others was found

One very pertinent PR relating to this: #1824

Yeah, I think your suggestion to return mktemp in the failure case is the right one.

I have a very vague recollection of an issue reported once by a user that
did not have a writable filesystem available. Are we undoing a fix here?

@WeatherGod: #1824 is one such issue. I think what's happened here is I broke support for that when I moved to using XDG directories on Linux. It looks, on further reflection, like we need to go the longer route and handle the case where get_home returns None everywhere it is called.