I've found a workaround to make script exit on sigint. Backend's event loop seems to block (sigmask or sigsuspend?) signals while waiting for a new event so we need to introduce an event that will fire without any interation with the window -- a timer. Next, all interrupts are catched somewhere inside matplotlib code so we need to make own SIGINT handler that will end the current process.
Here is the example code:

from matplotlib import pyplot as plt
import signal
import os

def sig(a, b):
    print("got sigint, exitting!")
    os._exit(0)

def timercb(e):
    print("timer")
signal.signal(signal.SIGINT, sig)

plt.plot(1)
fig, ax = plt.subplots()
timer = fig.canvas.new_timer(interval=5000)
timer.add_callback(timercb, ax)
timer.start()
plt.show()

I set the timer interval to 5 sec to make it sensible that signal arrives only after timer event happens.

attn @mdehoon

This is an interesting idea. I wonder if this approach (the timer, that is)
might be a suitable trick to deal with the other issues we are having with
the macosx backend and its eventloop?

On Sat, Jan 10, 2015 at 9:41 AM, Thomas A Caswell notifications@github.com
wrote:

attn @mdehoon https://github.com/mdehoon

—
Reply to this email directly or view it on GitHub
#3991 (comment)
.

@WeatherGod Which other issues are there with the macosx backend and its event loop?

I assume @WeatherGod is referring to bilt not working.

@WeatherGod

This is an interesting idea. I wonder if this approach (the timer, that is)
might be a suitable trick to deal with the other issues we are having with
the macosx backend and its eventloop?

When working on event loops, one has to resist the urge to use tricks to solve issues. Tricks may solve a particular symptom but you will end up with a bunch of hacks that will be very difficult to disentangle.

@tacaswell The bug in that case is not caused by the MacOSX backend, but by the upstream code trying to draw stuff outside of the event loop.

@mdehoon, Fair enough. and I can see a bunch of problems with the idea for
blitting anyway. In any case, what do you think of this PR?

On Mon, Jan 12, 2015 at 9:28 PM, mdehoon notifications@github.com wrote:

@tacaswell https://github.com/tacaswell The bug in that case is not
caused by the MacOSX backend, but by the upstream code trying to draw stuff
outside of the event loop.

—
Reply to this email directly or view it on GitHub
#3991 (comment)
.

@WeatherGod The suggested bug fix should not be applied, as this is not a bug in the MacOSX backend, but in Python itself. See issue 3180 on the Python website:
http://bugs.python.org/issue3180
Note that the bug also appears with other backends (e.g. Tkinter, see the example on the Python bug report).
It could help if you guys poke the Python developers on this issue if it stalls.
Depending on how issue 3180 is fixed by the Python developers, some changes to the MacOSX backend may be needed at that time. Let's keep this issue open until then.

This pull request #4006 allows the delivery of interrupts once Python is fixed.

Can someone on OSX try to reproduce using the snippet in OP?

I can still reproduce the original problem (Ctl-C does not close the window) with master and python 3.5.2 from Anaconda.

https://bugs.python.org/issue23237 <- @mdehoon 's upstream patch is still not merged, closing as there is nothing we can do about this.

@efiring does this problem also persist with mpl 2.0.0rc2 + IPython?

I can't run it from ipython on my system; I am using Anaconda, which is running ipython with a non-framework build of python. I can only test it as a standalone script using pythonw.