Great. Does this obsolete#940? One last thing I'd like to see before merging this is a short description in the developer documentation about a) how to run the tests with tox locally, b) how to get the results from Travis and c) the basics of how to participate in Travis, understanding that the last should be primarily a link to documentation elsewhere.

Looks great. Any draw backs to the service? I definitely think we should give this a trial run on master.

IIRC, Travis CI doesn't support GitHub organizations for enabling repos in the Travis UI, so you might run into some trouble following their instructions athttp://about.travis-ci.org/docs/user/getting-started/. IIRC, you can get around this by going to the repository in GitHub and going toAdmin | Service Hooks, and manually pasting in a personal GitHub account and Travis CI token.

This documentation looks great. Just ran your instructions for running tox in a clean virtualenv and all works great. I think this is going to be very handy. I'll wait 24 hours to see if there's further comments, and barring that, I think this should be merged and have the service hook set up.

One thing that would be nice to add to the documentation about tox.

"Testing all 4 versions of Python (2.6, 2.7, 3.1, and 3.2) requires having four versions of Python installed on your system and on the PATH. Depending on your operating system, you may want to use your package manager (such as apt-get, yum or MacPorts) to do this, or usepythonbrew <https://github.com/utahta/pythonbrew>_"

Is this merge-ready now?

Whilst I haven't run this branch myself, the changes look sound, so it has my +1.

I'm going to go ahead and merge this. Thanks for this, I think it is very handy.

Looks like you haven't gotten around to merging this yet?

Did you enablehttp://travis-ci.org/#!/matplotlib/matplotlib ?

Don't think I have permissions to do this. I was looking for ahttps://github.com/matplotlib/matplotlib/admin/hooks as per the guide athttp://about.travis-ci.org/docs/user/how-to-setup-and-trigger-the-hook-manually/.

@mdboom: do you have admin privs for the mpl repo or organization?

Didhttp://travis-ci.org/#!/matplotlib/matplotlib ever get set up?

@jdh2358: John, as I know you will have suitable permissions, please can you enable the travis hook on the matplotlib repository. Thanks!

@mdboom: We still need to set this up (it would have caught the tests failing). Are there other people with admin rights other than yourself?

I'm looking into this now. I have the service hook set up, and I have flipped the switch on the matplotlib/matplotlib repo on Travis, but I'm still not having much luck getting the build to trigger. Nowtravis-ci.org seems to be non-responsive.

On 2012/08/29 4:43 AM, Michael Droettboom wrote:

I'm looking into this now. I have the service hook set up, and I have
flipped the switch on the matplotlib/matplotlib repo on Travis, but I'm
still not having much luck getting the build to trigger. Now
|travis-ci.org| seems to be non-responsive.

—
Reply to this email directly or view it on GitHub
#948 (comment).

It is building, but not successfully. Among other problems, it looks
like it is starting by trying to build numpy, and going through an
excruciatingly long process.

Numpy takes a long time to build -- should probably have it install a pre-built numpy withapt-get oreasy_install. Let me know if you need help.

Can apt-get run on the travis buildbot?easy_install on a Linux box installs Numpy from source anyway, so that provides no real savings. (I know next to nothing about Travis).

Yeah, there are definitely.travis.ymls that invokeapt-get and it's officially supported (try searchinghttp://about.travis-ci.org/docs/user/build-configuration/ for "apt-get"). The question is whether theapt-get repo has a new enough version of numpy to use matplotlib and test everything that you want to test. If not, I guess you might be able to find an alternative repo or create your own. Here's an example.travis.yml that usesapt-get to install numpy and also useseasy_install to install nibabel. (it also does this weird trick where it pretends to be an erlang project rather than python so that they can create a custom python environment by hand -- you may or may not need to go that far -- I forget the exact details of when it's needed):

https://github.com/matthew-brett/nipy/blob/master/.travis.yml

I found a way to bypass the numpy compilation process. I set up a travis worker box and built binary dists of numpy for our testing environments (2.6, 2.7, 3.1, 3.2). The install section of travis.yml can now download and unzip numpy instead of compiling it.

install:  - PY_VER=`python -c "import sys;print('%d.%d'%sys.version_info[:2])"`  - wget https://github.com/downloads/pwuertz/travis-tests/travis_py${PY_VER}_numpy.tar.gz  - sudo tar -C / -xf travis_py${PY_VER}_numpy.tar.gz

Setup and the travis testing results:
https://github.com/pwuertz/travis-tests
http://travis-ci.org/#!/pwuertz/travis-tests

Its a nice solution, but it sounds a bit flakey. I'm assuming we can't just use a pre-built versions of numpy from official sources? This is bound to be a common problem for other projects depending on numpy. Maybe its worth spinning up a whole repo which has pre-built travis compatibletar.gz files (matplotlib also being one for use with upstream projects) which has a stable process for creating these tar files in the first place (via a.travis-ci.yml file perhaps).

No, there are no pre-built versions for linux. I think ideally one would just use a travis worker to build all dependencies required for testing matplotlib, build binary dists taylored to the travis environment and upload these somewhere. Travis announced support for building artifacts which could aid this process, but right now there is no official way of getting data out of the worker processes.

Ok, I just talked to the guys in the irc chat. There is currently no way of securely uploading data from a worker process. So setting up a VM like I did for doing the work is apparently the only way at the moment.