We might want to get rid of this script altogether in a follow-up PR. It is a lot more "trivial" now. Might make our CI more readable to not have the extra layer of abstraction.

OK, I got rid of it

Do we need Cython during testing?

Still curious about this. But not critical I think since we plan to remove this eventually anyway.

Yes, we should get rid of this as soon as free-threaded builds are more commonly available.

Should we make sure that $REQUIREMENT_FILES includes default.txt, and optional.txt (if needed) regardless of setup_test_env.sh being run? Also are we sure that $REQUIREMENT_FILES sticks from running setup_test_env.sh before?

This is a remnant from the previous script; I will remove $REQUIREMENT_FILES, since there's really no need to robustly check the docs build under minimum AND latest dependencies. User should only need to be able to build docs after pip install -r requirements/docs.txt.

Docs are a small subset of functionality that should be covered by regular test suite.
Docs examples should work, of course, which is why it needs to be run once. But library functionality should be tested in the test suite, and all platforms.

Hmm, might be good to check that there's a working dependency solution between "default", "test", "optional" and "docs" for local development. But agreed not critical right now. Might be something we think about, when / if we update our dependency groups though.

We could build some checking logic into tools/generate_requirements.py?

I think this is currently the only place in our CI where we fail on warnings for Sphinx. We should make sure to add that flag in another job.

To be honest, I would prefer if we discuss and deal with removing docs on Azure separately. See also #7658 that proposes a different route.

To be clear. Not a blocker, just noting that I'd like to keep testing our doc pipeline on Windows in some form. Might be good to get those faster runners from GitHub so this is no longer a huge time sink!

I emailed GH about the faster runners earlier today, and will be happy to reinstate the Windows docs build if we are able to get those going. The 4x build time is a major blocker right now for getting CI build times down. I saw the differently proposed route, and will be happy to add that if you and @jarrodmillman come to an agreement.

Noted about the warnings, thanks.

This could also sed the dependencies in pyproject.toml. Otherwise pip install . to be safe?

It requires a bit more finesse to do this in the pyproject.toml, because you'd need to avoid areas other than the test dependencies. I can double check that all pip installs are with --build-isolation.

This will miss the pytest config in pyproject.toml but we will deal with that in a separate PR.

Did you chose aarch64 / ARM because its the fastest right now?

Exactly; we can reduce the 40 minute doc build down to 10 minutes.

I'm wondering now why we actually use --no-build-isolation at all. Seems to me that we could remove the "Install build dependencies" step before and simply rely on pip install to do its thing?

Might be something we could look into as a follow-up simplification. But not terribly important.

You actually set it up this way already for INSTALL_FROM_SDIST which skips installing the builds and uses an isolated build env.

I suppose it shortens the subsequent testing and documentation dependency installation steps, but it's not a big deal; we could, indeed, remove that step.