More robust version checking

[larrybradley]

This PR uses LooseVersion for more robust and simpler version checking.

[bsipocz]

👍
thanks @larrybradley for crossing off things from my todo list :)

[bsipocz]

@larrybradley - The travis failure is totally unrelated and due to a conda issue I'm fixing in ci-helpers atm. When that one is in I meant to merge this.

[larrybradley]

@bsipocz Thanks, I was just trying to figure out what was going on with that test (it's failing on my other PR too). I thought it was a transient failure, but I quickly discovered that it wasn't.

[bsipocz]

@larrybradley - Hmm, it wasn't the silly install WARNING that got picked up. The only remaining warning is the matplolib one https://travis-ci.org/astropy/photutils/jobs/101686421#L521 but those usually don't trigger sphinx fails.

[larrybradley]

I have no idea why that is triggering the failure either. I suspect the matplotlib font cache warning is due to matplotlib 1.5.1 which was released two days ago, but I don't know how to fix it.

[bsipocz]

What about trying to fix the version number to 1.5.0 for now?

[larrybradley]

Thanks, I'll try that.

[astrofrog]

Had the same issue in Glue, need to fix Matplotlib to 1.5.0 - see glue-viz/glue#821 (comment)

[astrofrog]

But it's not a long term solution :(

[larrybradley]

Thanks, @astrofrog. Hopefully they fix this issue in the next matplotlib release.

[larrybradley]

@bsipocz Why is the build_sphinx test using Python 3.4 instead of 3.5? Does 3.5 not work?

[bsipocz]

@larrybradley - I don't find any trace of a reason for it, most probably it was as overlook.

[larrybradley]

matplotlib 1.5.0 conflicts with python 3.3 in conda (https://travis-ci.org/astropy/photutils/jobs/101925006#L419). So, I'll have to use pip for matplotlib.

[larrybradley]

Is there a way to specify a maximum version in conda (like conda install matplotlib<=1.5.0, which doesn't work)?

[bsipocz]

@larrybradley - Let's see how much slower the pip solution is. If significant, then I suggest to fix the version number only for the sphinx builds (as those are the only affected builds) and use the latest from conda for the rest.
Not sure that one can say the maximum version number, but what about saying matplotlib=1.5.0? It shouldn't install a newer version, I think.

[larrybradley]

I tried matplotlib=1.5.0 but the install failed because of a conflict with python 3.3:
https://travis-ci.org/astropy/photutils/jobs/101925006#L419

It appears matplotlib 1.5.0 is no longer available with python 3.3.

[bsipocz]

Sorry, what I meant was to have this:

        # Check for sphinx doc build warnings - we do this first because it
        # may run for a long time
        - python: 2.7
          env: ASTROPY_VERSION=development SETUP_CMD='build_sphinx -w'
                  CONDA_DEPENDENCIES='Cython scipy scikit-image matplotlib=1.5.0'
        - python: 3.4
          env: ASTROPY_VERSION=development SETUP_CMD='build_sphinx -w'
                  CONDA_DEPENDENCIES='Cython scipy scikit-image matplotlib=1.5.0'

[larrybradley]

Ah, yes that's probably what I should have done at the start instead of changing the global matplotlib dependency. Thanks! 😄

[larrybradley]

@bsipocz I tried fixing matplotlib=1.5.0 for only the Sphinx builds. matplotlib 1.5.0 does get installed, but then it gets updated to 1.5.1 (see https://travis-ci.org/astropy/photutils/jobs/101941930#L495). Do you know what's going on? Feel free to open a new PR if you can fix matplotlib to 1.5.0. You may have better luck than me. 😃

[larrybradley]

The fix is now upstream in ci-helpers (astropy/ci-helpers#41).

[larrybradley]

I'm going to merge this. The test failure is due to an errant warning being issued by matplotlib 1.5.1 -- it will fail until the matplotlib version is fixed to 1.5.0 in ci-helpers.