@mdboom - just wanted to ping you on this because I've changed the behaviour such that by default you get a white background for the figure (as per savefig) rather than adhering to the patch's facecolor.

I've taken liberties with these enhancements against v1.4.1 purely because the nbagg backend is pretty experimental and these changes make the backend more user friendly. I'd be pretty disappointed to put them against v1.5.x/master, but it wouldn't take much argument to encourage me to do it, if anybody has major concerns.

Notice that I've added a UAT document - automated testing of the nbagg backend is hard. In the future I can imagine some limited mock unit-tests, but there will always be a need for a few integration tests. I was torn between putting the UAT in the backend code vs the testing folder - I can live comfortably with either, so again, if anybody has a preference, just let me know.

@mdboom - just wanted to ping you on this because I've changed the behaviour such that by default you get a white background for the figure (as per savefig) rather than adhering to the patch's facecolor.

Scratch that - I've changed my mind and decided that simply adding a rcParam which determines if the nbagg defaults to transparent is the easiest and most consistent approach.

Looks good. I don't have a strong argument about where the UAT should live, but maybe it deserves a mention in the developer docs about running the tests?

Is there a good reason why the rcParam figure.transparent couldn't be reused here?

Is there a good reason why the rcParam figure.transparent couldn't be reused here?

I don't believe that exists currently? :

>>> print(matplotlib.rcParams.find_all('transparent'))
nbagg.transparent: True
savefig.transparent: False

The main reason for that specific commit is to change the default from a gray figure patch to something more useful (either white or, better still, transparent) when using the nbagg backend. I didn't want to venture into the realm of changing defaults for all backends, but feel that the nbagg backend has an opportunity to avoid ugly gray edges from the outset.

savefig.transparent, not figure.transparent is what I meant. Sorry for the confusion. And I see what you're saying -- don't want to change the defaults for everything at this point.

Looks like the web agg framework is struggling with transparency actually - I've only just seen it, but I'd say that the renderer isn't being cleared properly between draws, and therefore is sending over images with ghosting type effects. I'll try to take a closer look soon, but I suspect the quick solution is to set the figure facecolor to white for now.

Looks like the web agg framework is struggling with transparency actually - I've only just seen it, but I'd say that the renderer isn't being cleared properly between draws

That was exactly it. I've now fixed the webagg backend to solve this issue, and as a result, the nbagg backend is now functioning as expected. I'm very happy with the result of this PR - it turns nbagg into a genuinely usable backend.

P.S. I've tested this with both Python 2.7 and Python 3.4 - both work as expected.

Fixed pep8 issue locally, merged and pushed.

👍 Great stuff.

👍 Great stuff.

Except I broke the image diff updates - #3567 addresses the problem.