setuptool's setup_requires sandboxing breaks Cython

[ghost]

setuptool's setup_requires's easy_install sandboxing appears to break Cython/.pxd cimport targets by how it loads or unloads modules.

Sadly, build-system.requires is not a good alternative for Cython build deps specifically because using it for this can slow down metadata analysis incredibly, as I explained here: #1742 (comment) which is why setup_requires was also suggested to me as possibly better option here: https://discuss.python.org/t/support-for-build-and-run-time-dependencies/1513/61?u=jtt

The Cython ticket where the unfortunate easy_install interaction was observed is here: cython/cython#2730 (comment)

In overall this greatly affects my build time negatively which just adds up over a while, so it would be nice if there could be found some way to address this. I'd also be happy to migrate to some alternative to setup_requires, in fact right now I am using build-system.requires, but right now there doesn't really appear to be a good one for the reasons explained/referenced above which is why I'm hoping for a fix

[pradyunsg]

This is blocked on #1742, since it does not make sense to improve upon deprecated functionality.

I say let's determine further action there and then come to this issue. (obviously defering to setuptools maintainers if they prefer some other approach)

[pganssle]

We're probably not going to be making any improvements to easy_install, as we don't have the resources to maintain a parallel infrastructure for installation, and there's really no reason to do so, anyway.

I think that build-system.requires is the way to go here, and any problems with build speed need to be addressed in pip or PEP 517/518 (I also very much would like to see an option for incremental builds and build caching, for example).

[ghost]

@pganssle build-system.requires has the wrong semantics though, see in #1742 (comment) (and maybe it would also make more sense to discuss this over there, sorry for my slight tickets mess) - so I'm pretty sure it cannot be fixed with caching or anything in the implementation, although an entirely new option would of course be feasible

[jaraco]

As mentioned above, we won't be investing effort in advancing easy-install-based functionality. Please see #2823 where that functionality is being removed.