Re: [PATCH] import: pypi: read requirements from wheels.

[Ludovic Courtès]

Cyril Roelandt <address@hidden> skribis:

> On 01/24/2016 09:08 PM, Ludovic Courtès wrote:
>> Cyril Roelandt <address@hidden> skribis:
>> 
>>> * guix/import/pypi.scm (latest-wheel-release): New function.
>> 
>> s/function/procedure/  :-)
>> 
>> Please also mention the changes in ‘guess-requirements’,
>> ‘compute-inputs’, etc.
>> 
>> So do I get it right that pypi now provides packages both in Wheels and
>> in “traditional” format, but that Wheels provides more info about
>> dependencies?
>> 
>> IOW: What does this buy us?  :-)
>
> When uploading a package to PyPI, one may provide a wheel in addition to
> a tarball. The wheel is built by setuptools, and contains metadata,
> which includes *all* the requirements specified by the authors.
>
> It is much better than reading "requirements.txt", because this file is
> not mandatory.
>
> Since wheels may not be provided, we need to keep the old way of
> retrieving dependencies, even though it is not extremely reliable.

OK, got it, that’s nice.  Could you add a comment at a relevant place in
the code to explain that?

>>> +      (and (system* "unzip" "-q" wheel-archive json-file)
>>> +           (dynamic-wind
>>> +             (const #t)
>>> +             (lambda ()
>>> +               (call-with-input-file json-file
>>> +                 (lambda (port)
>>> +                   (let* ((metadata (json->scm port))
>>> +                          (run_requires (hash-ref metadata "run_requires"))
>>> +                          (requirements (hash-ref (list-ref run_requires 0)
>>> +                                                  "requires")))
>>> +                     (map (lambda (r)
>>> +                            (python->package-name (clean-requirement r)))
>>> +                          requirements)))))
>>> +             (lambda ()
>>> +               (delete-file json-file)
>>> +               (rmdir dirname))))))
>> 
>> Eventually I wonder if we should do this in a derivation instead of
>> hoping for ‘unzip’ & co. to be available there (“eventually”, because
>> the problem is already present with the ‘requirements.txt’ thingie.)
>> 
>
> I think it is fine as is, seeing how it makes Python packaging much
> easier :)

I mean the functionality itself is perfect, but what I mean is that the
above process (taking an archive + unzip + guile-json as input,
producing a list of requirements as an output) could be modeled as a
derivation.  Anyway, no rush for that.

Thanks,
Ludo’.