Closed
Description
I'd rather mirror closely what numpy and scipy are doing, while we are getting more familiar with the release process.
I guess it was done to be more convenient, you build and test all the wheels, approve upload on testpypi.org, test locally, and then can approve the upload to PyPI. Without the chain workflow you have to build and test twice, which is quite costly since right now can take 1-2 hours (because of the macOS 5 concurent runners limitation + the fact that we have 7 slow macOS-intel build and 7 faster macOS arm builds).
Metadata
Metadata
Assignees
Labels
No labels
Activity
jeremiedbb commentedon Sep 16, 2026
I would not care too much about following numpy closely. We have different release workflows because we have different constraints, and it's fine.
In #16 I'm reworking the workflows and now you have to run the publish workflow twice: first for testpypi, then for pypi. But each run can reuse the wheels built by the
build-wheelsworkflow.I thought about a publish workflow that would do both in a row, but I felt that it would not feel great if for example: testpypi passes then pypi fails; you have to fix and "rerun from failed". I found it more natural to fix and rerun the workflow for pypi instead.
Also it's a lot closer than what we used to do in sklearn so far with the
publish_pypiworkflow.jeremiedbb commentedon Sep 17, 2026
I also opened #18 as an alternative to avoid having to look for artifacts from a previous workflow. In this PR it chains build, upload to PyPI ans upload to TestPyPI. So it acts as the opposite of what you're suggesting here. I prefer this way because it guarantees that we publish the same thing on testpypi and pypi, and avoids paying the build twice.