Test Numpy on Python 2 / 3 on all architectures

[jgavris]

I'm working on getting numpy built for Python 3.6 on x86_64 and arm64-v8a ABIs, I just wanted to add and or check if this passes on CI in the meantime.

The test matrix could probably be refactored to be a bit cleaner, I think it would be great if every build command were tested against every ABI.

[AndreMiras]

Nice, thanks for the pull request.
I agree we should have a better coverage, but Jenkins would never be ending for every single PR.
So we need to first tackle this story #1382
Basically the idea is to dynamically build only the part that were touched by code change. That way we could cover more (e.g. different Python host/target version and architecture) the concerned recipes.
So I'm keeping your PR open, but can't merge it before #1382 because it would really cause too much overhead on the CI. Once we've done with it we will review/rework your PR to make it to the tree.
Does it make sense to you?

[jgavris]

What are the expensive parts of the docker build? I would be concerned it is difficult to get 'conditional builds' done correctly (without going all-in with a fancy build system like bazel). I wonder if there is any low-hanging fruit like rebuilding the docker images with up to date system packages and android SDK / NDK. That looks like it's taking a real long time, versus the actual python-for-android apk build.

[AndreMiras]

Yes the docker build takes time since it needs to download some heavy component for instance.
But not only, also compiling p4a recipes take time.
You can take a look at the last master build for instance:
https://travis-ci.org/kivy/python-for-android/builds/435279943
And open some of the different jobs to see where the time is spent. For instance in Job #173.1. Here docker build --tag=p4a . 1000 seconds (16 minutes) are spent and the rest (10 minutes) is spent on the docker run that builds the recipes.

So yes SDK/NDK takes ages, but it's not recommended to cache it in Travis. Yes one solution could be to use a Docker image that has SDK/NDK and is optimized, but I also like that we have the control over our Docker image.

Yes the conditional build is not an easy one, but I still imagine it's feasible.

[AndreMiras]

Thank you again for your PR! I'm sorry we didn't make it to the tree.
The support for x86_64 architecture was added recently in #1476.

Conditional builds are in place via:

• https://github.com/kivy/python-for-android/blob/12ac837/ci/rebuild_updated_recipes.py
• https://github.com/kivy/python-for-android/blob/0c936d1/.travis.yml#L27

It's still far from being perfect, so feel free to open a dedicated PR/issue if you want to further discuss the CI improvements. For instance yes pre-built docker image may help. Also the recent drop of CrystaX support in the CI speed up the Docker build.