Trial GitHub Actions vs Cirrus CI

[trexfeathers]

📰 Custom Issue

@ocefpaf has correctly pointed out (SciTools/cf-units#175, SciTools/nc-time-axis#66) that GHA is the most popular CI implementation for scientific Python projects. Given that barriers to involvement are a big factor in the health of a project: we need a really good reason to go against the crowd as we do now using Cirrus CI.

I believe our reasons for using Cirrus are to minimise the runtime for a full CI run. We should therefore do a trial to see if it possible to get similar/faster runtime using GHA instead. I'm told that GHA also has caching solutions like Cirrus does, and the minutes are free for public repositories so there should be no trouble with parallelism.

[trexfeathers]

I probably shouldn't have been so quick to shoot down @jamesp's idea on SciTools/iris-grib#242!

[jamesp]

@trexfeathers it wasn't the right solution at the time, it might be now... either way all good!

[rcomer]

Cartopy went to GHA in December SciTools/cartopy#1691

[bjlittle]

We could easily move to GH actions for CI. I'm certainly not wedded to cirrus-ci - it certainly dug us out of a hole quickly when travis-ci was causing us major issues. Also, what's seriously going against cirrus-ci is having to pay credits for reasonable resource 🧐

I've seen several major projects migrating and using it, and I'm bootstrapping GeoVista with GH actions instead of cirrus-ci. It has an interesting use case, as it also needs Mesa support for PyVista, and I want to target linux, osx and eventually windows (but iris needs a bit of work for that) - anyways, all of that is easy to do and supported for GH actions.

I don't believe this will be a difficult thing to do, and the interface with nox as it stands, shouldn't need any, or very little, change.

[bjlittle]

Also, if we're rebooting CI, for me idealistically this would include:

• testing on ubuntu-latest, osx-latest, windows-latest
• testing both conda and pypi packaging
• pushing a slim / distro-less docker image to dockerhub for stable releases

[bjlittle]

We should openly discuss nox vs tox

[bjlittle]

Also see SciTools/nc-time-axis#57

[jamesp]

Agree it's worth considering tox vs nox. See https://github.com/tox-dev/tox-conda which has support for @EXPLICIT conda-lock files

[trexfeathers]

It sounds like there are some exciting possibilities for parallelisation with GHA+pytest, so that might be another argument for switching.

[tkknight]

Worth considering using GH actions to build the docs also on each pull request so that it is viewable by the reviewers.

Discussion: #4398

[pp-mo]

#4503 probably addresses