[RFC] Improve Python package quality

[Lars-Dominik Braun]

Hi,

I’d like to propose adding an additional phase to python-build-system to
improve Guix’ Python package quality.

Python is an interpreted language without any compile-time checks. Any
errors are only visible at run-time, including missing or wrong imports.
Additionally most Python packages use an additional packaging layer
through setuptools, which also declares mandatory and optional
dependencies.

This can result in buildable, but broken packages. An example of a
botched upgrade is fixed in commit
22e06297b1982f75aaadddba616b1052e506e4a0. The package python-httpx was
upgraded, but that new version declared a dependency to a newer
python-httpcore via setuptools, which wasn’t available and thus packages
depending on python-httpx were broken.

My proposal adds some build-time checks to guarentee three properties:

1) Depending on the package (distribution in Python language) via
setuptools works. This ensures dependencies are complete and have the
correct version as determined by setup.{py,cfg}.
2) Console entry points can be loaded. This ensures scripts installed to
bin/ actually work.
3) Top-level modules can be imported. This ensures the package is
actually usable and has no undeclared dependencies.

The attached patch implements all three. I’ve been rebuilding a few
Python packages using the patch and uncovered some issues already. I’m
sure it’s not perfect yet and thus I’m open to improvements. If this
idea is well received I’m willing to rebuild and fix *all* Python
packages broken by this new phase. I’m also aware this change needs to
go to core-updates.

Cheers,
Lars

0001-WIP-build-system-python-Validate-installed-package.patch
Description: Text Data