Hmm, maybe it makes more sense to switch it on the DNS side. I'll try that first.

I think this makes more sense as far as a test goes.

@astrofrog it appears to not have worked anyway.

Trying to check if wheels.astropy.org is pointing to the new URL on travis.

DO NOT MERGE!

@Cadair @eteq - ok, so this is the solution:

• DNS has been changed so primary server is mine, which is only updated manually
• Versions are all fixed now to ensure no unexpected breakage in future (most of them were already fixed except a few)

I will merge once this passes so that we can rebase the other pull requests.

@Cadair - I've restarted the following build which should be using the previous settings, and I think it will use your wheels so we can see if it works:

https://travis-ci.org/astropy/astropy/builds/14702152

If it uses matplotlib 1.3.0, it's using my wheels, and if 1.3.1 then yours. If 1.3.1 passes it means your new wheels work.

@astrofrog - this is confusing me: the test you indicated in fact failed the sphinx build, but it seems to be using mpl 1.3.1 ... Is that what was supposed to happen, and the failure is unrelated?

@eteq - the problematic wheel is the one for matplotlib 1.3.1 that was made yesterday. The above build I pointed out shows that this wheel does not work. What I've done is to explicitly request 1.3.0 because that definitely worked, and I switched the DNS to my server to make sure we use the original version from my mirror is used. The other reason for switching to my mirror is to ensure stability, because @Cadair's gets automatically updated by Travis, so by using mine as the primary one, we have more control over updates.

What's weird is that this is not an issue with matplotlib 1.3.1 as such because it was working earlier this week, but somehow when the wheels got re-generated yesterday, something changed and the plot directive no longer works.

@astrofrog - Ah, I gotcha - so that ended up showing demonstrating was not working. That's what was confusing me - on the same page now.

I just tried building the docs now locally with mpl 1.3.1 and it worked just fine. Perhaps order matters? That is, maybe the plot_directive doesn't work if sphinx isn't installed prior to building matplotlib? @mdboom might be able to shed some light on this.

@eteq - good point, and it could also be to do with the order of the wheel generation?

Interesting, but neither of those have changed!!