I'm sure that's a solution but sure feels icky. Not that we shouldn't merge if we don't find anything better. Does putting inline imports in __del__ work or does importing from del introduce its own set of problems? Somehow that feels less likely to break in the future.

I remember a problem like this from another project, but I don't really remember what was the correct solution, reimport or holding on to a reference. I don't know or forgot how to provoke such errors though.

Maybe one could also tweak LatexManager a bit so it doesn't involve __del__ at all, that would probably be the cleanest and most pythonic solution. But then I would have to re-test the reliable deletion of temp files again and I don't have a windows system at disposal.

Is this a case where atexit functions would be useful?

The problem is specifically with __del__ method in LatexManager. atexit is already used in the module but the __del__ method seems necessary for other times than shutdown.

As I initially reported, the problem that this patch deals with popped up in the project Sage. See http://trac.sagemath.org/ticket/14798 This ticket will be closed once any solution is merged to matplotlib.

If other solution is suggested, then I can test it with Sage. Also the problem is not only with Windows. My Sage installation is on Mac.

I am writing since I am wondering what would be done for this pull request, as I am ignorant of the general policies of matplotlib development. For this pull request to be accepted for a merge to the next release of matplotlib, should someone give a positive review or is this just up to the release manager?

The solution in this pull request works though I am not completely satisfied with its method. So I would be happy if this pull request is accepted, unless someone proposes a better solution.

Yeah -- we should probably hold our noses and merge as-is. 😉

Backported to 1.5.x as a561593