So… It seems that USE_PYTEST == false and true do not have the same behaviour.
I am investigating…

Selfishly I am 👍 for CI build of the docs on Python 2.7 😈. FWIW, I have just successfully locally built the doc on Python 2.7 (and Linux), with #8165 in its current state (eventhough I still have to have a look to a comment @anntzer did on this PR).

I would rather a 3.5 test run than building the docs twice, the docs are just a smoke test (we should improve the test coverage of the validators).

Appveyor covers 3.5.

@afvincent indeed your patch will fix the documentation under 2.7.

@afvincent @NelleV There's a bunch of other entries in rcsetup.py that check for six.text_types. I have a strong suspicion that most of them should be replaced by six.string_types, otherwise they'll just break in the same way when a non-unicode string is passed in Py2.

Yep, that's the conclusion I came up with.

I would also be inclined to agree with this conclusion: in #8165 I used six.text_type in first place because it seemed to be used quite a lot to do this kind of checking in rcsetup. And I was bitten rather bad for doing so instead of using six.text_type… However, based on my further experience in #8165, I would advise to be careful with six.string_types too, as it would also let pass strings like "Hello".encode("utf-16") (or with any other weird encoding that does not map smoothly with the ASCII one), which can have some weird/funny/unexpected side-effects too.

@QuLogic It appears that on a successful build, Appveyor will create a wheel to upload to conda (see appveyor.yml after_test). Perhaps we could save some time by always first building the wheel and then installing the wheel to run the tests? (especially if we don't have a C compiler cache)

@anntzer That sounds like a good idea; I'm not sure why it was done this way or why we need to build the wheels every time either. Ping @JanSchulz.

The wheels are not uploaded, but just placed in the appveyor download "spot". The idea was to have a place where win users can download stuff to try fixes.

Sure, but they could still be built first and the test installation done using them.

Quick, related questions: to me it appears that we always use python 2.7 to build the docs, no matter which version of python we've set travis to use: https://github.com/matplotlib/matplotlib/blob/master/ci/travis/test_script.sh#L38

Is that true or am I missing something?

That's just to run linkchecker.

Ahh yes, i see it now.

(But in fact linkcheck is now builtin to sphinx so we don't even need that. See e.g. #5380.)

Closing this as we are now down to:

• on circleci
  • 3.5 docs
  • 2.7 docs
• on travis
  • Linux 2.7
  • Linux 3.4
  • Linux 3.6
  • OSX 2.7
• on appveyor
  • 3.5
  • 3.6

We need to try turning the 2.7 build back on on appveyor and I might try adding a 3.6 build on OSX.