On Monday, June 2, 2014 4:05:03 PM UTC-7, Stéfan van der Walt wrote:
On Tue, Jun 3, 2014 at 1:01 AM, Michael Aye <kmicha...@gmail.com> wrote:
> I just went through a quite painful learning process with all this, because
> the docs are quite scattered, with lots of old stuff mixed in. So I'd like
> to comment, in case you don't insist on the Travis build machinery to be
> MacPython (is that even possible, as Travis runs on Linux?), to speed things
> up IMMENSELY over using pip for requirements install, one can use conda for
> the env setup in Travis. It speeds things up, because pip doesn't store
> binaries and instead spends a lot of time compiling things (or so it says
> here: http://sburns.org/2014/03/28/faster-travis-builds.html). Find the
> instructions on how to setup Travis with conda in that link, I tried it and
> it works fine here.
> I did one thing different compared to that blog entry though: In the travis
> yaml I prefer to "script: python setup.py develop && python setup.py test",
> because that way I don't get a useless test run in case the develop install
> fails due to a bug.

This works great--I've used that approach for our Windows buildbot too.

On Linux, since we have apt-get, it's less of an issue, even though we
currently build Matplotlib from source to get the latest version.

Some of us are building wheels for the travis linux machines too - see for example:

https://github.com/matthew-brett/nipy/blob/12f96d576be0b7c841a65b095ae0b70a22b36d2b/.travis.yml#L20

That gives me the latest released version of matplotlib.  The wheels are pretty easy to build too, just create a vagrant virtual machine, and you're ready to go:

https://gist.github.com/matthew-brett/714b50bd3159d416981a

Cheers,

Matthew