Update: This only happens when using the (default) osx backend.

Are you using the framework build of python? (ie. are you starting python using pythonw)

No. It works with pythonw, but how can I start IPython using this framework build?

Here is a magic incantation to get ipython to play w pythonw #8982 (comment)

Nice, thanks @jklymak this works! I wish Anaconda made pythonw the default for everything to avoid these kinds of tricks. Has anyone ever asked them why they don't want to play nice with macOS frameworks?

Ha ha. I think there are a few threads about it.... I don't understand the issue, but, I don't use ipython except via jupyter notebooks, so...

OMG what's the tl;dr? I assume not much has happened. This is the one reason why I can't recommend Anaconda on macOS without any reservations.

There is an open issue: ContinuumIO/anaconda-issues#199

if you want more bedtime reading ;-)

You're right, but the problem with this is that osx is the default backend. If it wasn't this wouldn't be such a big issue. Anaconda could probably change the default to qt5 because PyQt5 gets installed anyway.

#9795 will make it easier to configure the "default" on a user-by-user basis...

I thought qt5 was the default on osx with conda....Did that change recently?

[I just checked by installing from defaults (version 2.0.2) and from conda-forge (version 2.1.0), and in both cases the backend default is qt5agg.] This is with Python 3.6.
Correction: The original comment above (now in brackets) is incorrect. Both versions provide matplotlibrc files in which the backend is macosx. The conda recipes are not modifying the default that we supply.

How did you check this? When I start IPython in Anaconda, I get:

In [1]: %matplotlib
Using matplotlib backend: MacOSX

Also, matplotlib.get_backend() returns MacOSX.

@cbrnr, thanks for keeping after this--I wasn't being careful enough. It turns out I didn't check something that I thought I had checked, and the upshot is that my backend was being set from my own matplotlibrc file, not from the one installed by conda.

Side note: it's far from obvious, but to find out which matplotlibrc file is being used one can call matplotlib.matplotlib_fname().

So what's the plan? The best solution would be if Anaconda defaulted their Python executables to a framework build (something everyone else is doing, seems to have no negative side effects and makes Python run smoothly on macOS). Alternatively, as a workaround, they could change the default matplotlib backend to qt5 (on Macs) - not sure if this is feasible though because this would mean introducing PyQt5 as a hard dependency of matplotlib.

@cbrnr, could you test whether the issue is gone with the fix #12633?

import matplotlib
matplotlib.use('macosx')
import matplotlib.pyplot as plt
plt.plot(1)
plt.show()

I tried, but cloning stalls after an hour or so, so I can't test with this PR. However, I noticed that my example from my initial post does work in Anaconda now. Unless I create a new env, I then get a segfault. So I'm not sure what's wrong with Anaconda on my machine, I never seem to get it to work properly.

@cbrnr, thanks! Keep us informed if possible.

I tested with the latest Anaconda packages and this issue has been resolved. Thanks!