Re: python importers as an alternative to propagated-inputs

[Ricardo Wurmus <rekado@elephly.net>]

Justin Veilleux <terramorpha@cock.li> writes:

> Hi everyone. I was thinking about the propagated-inputs field in package
> definitions. As I understand it, it is useful as a way to replace RPATHs
> in packages that aren't compiled or don't support them.
>
> I was reading the documentation on
> https://docs.python.org/3/reference/import.html. It looks like we can
> define custom objects to do module resolution, possibly bypassing
> PYTHON_PATH lookup. I think it would be possible to write a very simple
> importer object that looks up module paths from an environment variable
> that looks like this:
>
> PYTHON_GUIX_MODULE_PATH=numpy=/gnu/store/...,pandas=/gnu/store/...

How would these compose?  “pandas” has dependencies, too.  Where would
those be read from?  How can be guaranteed that these are compatible
with what we had at compile time?

> Does someone know if this has been tried before? What potential problems
> would I encounter if I tried to implement this?

In 2018 we discussed the problems of PYTHONPATH extensively.  Hartmut
Goebel provided a three part analysis of the problem and potential
fixes.

The discussion didn’t lead to a generic solution, but it provided the
minimal reasons to replace our use of PYTHONPATH with GUIX_PYTHONPATH,
which we enable with a generated “sitecustomize.py” file — see (gnu
packages python) for details.

It would be great if we could do without GUIX_PYTHONPATH, which forces
us to harmonize *all* Python packages whenever we upgrade any of them.
Something that doesn’t require a global search path at all would be
ideal.  Perhaps the venv feature is how we can get there.  I recommend
reading the “PYTHONPATH issue analysis” messages for ideas and a list of
open questions.

-- 
Ricardo