I'm pretty sure that comment from 2004 about anim.py actually predates the current animation module.

On OS X, the following

(test) efiring@manini2:~/work/programs/py/mpl/matplotlib/examples/animation$ ipython
Python 3.5.2 |Anaconda custom (x86_64)| (default, Jul  2 2016, 17:52:12)
Type 'copyright', 'credits' or 'license' for more information
IPython 6.2.1 -- An enhanced Interactive Python. Type '?' for help.

In [1]: import matplotlib

In [2]: matplotlib.use('tkagg')

In [3]: run simple_anim.py

hangs with a spinning ball when I close the window with the window manager button. Control-C ends it.
With qt5agg backend, the window closes on command.
It also closes on command with tkagg backend on master, so it looks like this PR might be fixing one problem and introducing another.

I see that this patch also adds an error message in the same context as given by @efiring but on Linux:

invalid command name "140521287480648_on_timer"
    while executing
"140521287480648_on_timer"
    ("after" script)

which is not solved by #9956 either.

We probably need to look into clearing the timers at destroy time too. Registering the Timers in a (weak)set and stopping them the the figuremanager's destroy hook doesn't seem to help.

Actually, this is the situation on Linux:

• Original: Keeping the call to update() and the slow interval on the example: no error from command line or ipython
• Keep the call to update() and switch to fast timer: I get the tk-level error ("invalid command name etc.") and the python traceback (both command line and IPython)
• Remove the call to update() and keep a slow timer: no error
• This PR: Remove the call to update and switch to a fast timer: I get the tk-level error only from ipython, no error at the command line.

So from this point of view the PR is strictly better, even though it doesn't fix all issues.

Perhaps I can just switch back the example to use an slow timer if the failure mode is more severe on OSX than on Linux (but the same as before the removal of the call to update()).

Some further research pointed to https://bugs.python.org/issue1252236 (by Matplotlib's own @mdehoon :-)) which indicates that tkinter only checks for events every 20ms (the line static int Tkinter_busywaitinterval = 20; is still present in the sources); and, indeed, it looks like the erroneous behavior starts to appear as the interval goes from above 20ms to below (I tried 25 and 15; at 15 you may need to try a few times before reproducing into the issue, probably because this presumably requires some whole interval to occur between two event checks, or something similar).

This doesn't mean that the issue is necessarily unsolvable, but may point out why certain attempts don't work...

Would be nice to get this in so that I can make a tkcairo PR that doesn't cause conflicts with this.