Do you have any idea what the test coverage for this change is like?

Sorry, my new hobby horse is test coverage after the boxplot refactor introduces ~5 major regressions.

@tacaswell perhaps we should consider setting up something like https://coveralls.io/ with the matplotlib travis to get a feel for the coverage?

I have looked at the coverage briefly and it is a depressing situation (like 60% depressing). Setting up coveralls is a fine idea (so long as we keep it from posting on every PR thread on every commit).

@tacaswell I fully support the coverage hobby horse in general ... the caveat being that I'm already busy beating coverage into submission on my main project :). Coverage monitoring from travis would be really sweet -- I would expect that part of travis' output should be if coverage decreases on a particular PR. I find that coverage's html output is plenty good enough for interactive inspection of it.

Speaking of this PR in particular: Both code hunks are covered in the test run on my linux machine. I couldn't speak intelligently as to whether there are multiple diverse states in which coverage is necessary here.

I just wanted to add that I've tested this against cartopy (which is a big user of the affine/non-affine separation) and there seems to be nothing out of the ordinary (cartopy tests don't actually pass with >=v1.4.0) therefore I'm 👍 on this.

Nice little optimisation!

@pelson What did we break that breaks the cartopy tests?

What did we break that breaks the cartopy tests?

Ah, we just moved a few clip areas - nothing major (and that wasn't picked up in mpl's tests) just not pixel-pixel perfect between v1.3 and v1.4.