Taking Github Actions for a walk in the park
Travis started to scare me a bit
People were talking about github actions for a while now, lot of hype around it. I myself was quite happy with the things I’ve managed to get done with Travis.
But then came in this announcement from travis: “The new pricing model for travis-ci.com”. I was reading via lot of the backlash in tweeter about it, and after carefully reading the announcement itself, I was starting to think that the work I did for scylla-driver was counting on Travis too much.
Recently I’ve start seeing long queues for job triggers in travis, and I was starting to have second thoughts about picking travis. (not so long ago, less than a year ago)
The sweet setup I had in Travis
As a recap for scylla-driver, since it’s a c based package, heavily using cython. We are building python wheels for windows, mac, linux, ranging from python2.7 to python3.9, and pypy. (also using the Travis experimental aarch64 and ppc64le support)
For building the wheels we are using cibuildwheel, which is a nifty tool that magically keep all know how, on how building wheels correctly on multiple platforms. In linux, it’s using the official docker images provided by the [Python Packaging Authority ] for manylinux2014. And on windows and mac equivalents with the correct set to needed tool to get you wheel happily containing all the need dynamic libraries it depended on
cibuildwheel also comes with readily available examples for all the major CI system out there
cibuildwheel is also helping you in running unittest ontop of each built wheel (since in my case the setup is a bit complex, getting it installed and compile across the board was hard, we we’re skipping some unittest on some platforms)
Travis was also running our integration test suite, and running it on our PRs, and automatically uploading all the wheels and source distribution to Pypi.
I was working on this setup for long week, with a tight deadline for it to be ready for europython2020, thanks to @ultrabug
Github actions
For a while I was a bit unsure if it would be a match to what I had in travis. But one day our own Jenkins server was down for maintenance, and a big portion of my day was clear case of that.
I’ve started with copy-pasting cibuildwheel github actions example, and worked my way from there.
When I’ve started adding all the different unittest and platforms, I was struggling a bit with how the matrix is working, and how conditional steps works.
It ended up something like this, almost identical to what I had in travis:
name: Build and upload to PyPi
on: [push, pull_request]
env:
...
jobs:
build_wheels:
name: Build wheels $ ($)
if: contains(github.event.pull_request.labels.*.name, 'test-build') || github.event_name == 'push' && endsWith(github.event.ref, 'scylla')
runs-on: $
strategy:
fail-fast: false
matrix:
include:
- os: ubuntu-18.04
platform: x86_64
- os: ubuntu-18.04
platform: i686
- os: ubuntu-18.04
platform: PyPy
- os: windows-latest
platform: win32
- os: windows-latest
platform: win64
- os: windows-latest
platform: PyPy
- os: macos-latest
platform: all
- os: macos-latest
platform: PyPy
We had a nice twist that we didn’t have before, but now we can control from github PR which part of my pipeline would run using label, since github actions is so tightly integrated you can lookup almost anything related to the PR.
Github Actions - UI
This is how the final thing looks like:
logging is a bit weird to watch while it’s running, but after the fact the log are very tidy to look at
Late surprise - cross compile using qemu
With almost perfect timing I’ve run into this https://ift.tt/380SGhY by Pavel Savchenko
I was trying all kind of things to get aarch64 support, but this PR figure it out quite nicely, with one magic step:
- name: Set up QEMU
id: qemu
uses: docker/setup-qemu-action@v1
with:
platforms: all
For reference the full workflow
Now we can have docker using qemu to run docker instances based on different cpu architecture, it’s horribly slow but it’s working.
And nice side effect I’ve learned how to run it locally, so I can debug those things. Something I haven’t figured when running with Travis, apparently travis was using the same trick on their experimental support for those architectures.
Summary
Github actions works out of the box, with almost zero fraction, almost anything I could think about I’ve found a readily available step I can pickup, very impressing.
One thing I haven’t yet tried out, is using self hosted runner, hopefully I won’t need it. Cause who want to babysit anther server on his own, or worse, a bunch of them.
from Planet Python
via read more
No comments:
Post a Comment