Is this based on a template taken from somewhere?

The actions are from another project I worked on with similar build requirements. The docker parts for the manylinux build come from the pypa demo project

https://github.com/pypa/python-manylinux-demo

You can see results of this in my own repo (in progress)

https://github.com/WillAyd/pandas-release/pull/1/checks?check_run_id=441681755

OK got a minimum example working on GitHub actions. Here is what it looks like:

https://github.com/WillAyd/pandas-release/pull/1/checks?check_run_id=441916246

The artifacts are really zip files containing individual wheels. We could alternately put all of the wheels into one zip file

A few things worth discussing:

1. GH Actions by default only has a Windows environment to build a win_amd64 wheel. We also produce win32 wheels today, which I think are 32 bit Windows wheels (not sure how Windows names things). If that is correct, would need another way to build 32 bit Windows (maybe Docker)
2. The macOS wheels come with platform tag macosx_10_13_x86_64 (the actual environment is 10.15). We offer 10_9 on PyPi, so if we want to go that far back again need another option. Note that 10_13 came out in Sep 2017 and 10_9 came out in Oct 2013
3. I didn't add a make rule to download / unpack the wheels, though I think we would need that

One other thing I had but remove from the process was installing and running pandas.test() as part of the action. This seemed to fail by default on macOS / Windows and the installation for the ubuntu images are slightly more complex, so didn't bother with them for now

Would love any feedback anyone has to offer

we wouldn’t need to actually run .test() because by definition we only would build these after the. CI is green (it would be nice to make sure that the wheels actually install properly)

we should just drop win32 entirely

Updated to drop 32 bit wheel generation. I also added a step to install and run import pandas with the wheel to be sure

My main concern is whether we want to start using something custom, versus tooling that is also used (and thus partly maintained) by other packages in the ecosystem (multibuild is now used by many like numpy, scipy, scikit-learn, ...)

Its a good question. I just think this is much easier than trying to configure multi build and keeps everything in Azure

FYI: MacPython/scikit-learn-wheels#23 (building on azure-pipelines)