[Buildroot] [git commit] package/python-setuptools: add missing dependency on host-python-wheel

[Arnout Vandecappelle]

On 03/07/2023 20:56, Thomas Petazzoni wrote:
> On Mon, 3 Jul 2023 20:20:31 +0200
> Arnout Vandecappelle via buildroot <buildroot at buildroot.org> wrote:
> 
>> commit: https://git.buildroot.net/buildroot/commit/?id=3b2913552ee4e304d4d86e577b237652a8f0a99f
>> branch: https://git.buildroot.net/buildroot/commit/?id=refs/heads/master
>>
>> While migrating to pep517 build backend the host-python-wheel
>> dependency has been added to host-python-setuptools but
>> not to python-setuptools. Note that host-python-wheel is not really
>> needed during the build of (target) setuptools, but the setup.py script
>> checks if the dependency is present in the host directory.
>>
>> To make everything consistent, add host-python-wheel as a build
>> dependency, and BR2_PACKAGE_PYTHON_WHEEL as a runtime dependency.
> 
> Hu, this was not in the patch from Romain I believe.
> 
>> +	select BR2_PACKAGE_PYTHON_WHEEL # runtime
> 
> This is not correct, there is no such symbol in Buildroot, and it
> causes a check-symbol error:
> 
> https://gitlab.com/buildroot.org/buildroot/-/jobs/4586144848
> 
> What was the motivation behind this additional select?

  Because setuptools has a runtime dependency on wheel. The pep517 
infrastructure does a build-time check of this runtime dependency, but due to 
the way we run python, the build-time check looks at what is installed for the 
host, not what is installed for the target. Therefore, we have a dependency on 
host-python-wheel instead of target python-wheel.

  Since setuptools clearly defines that it has a dependency on wheel, I thought 
it logical that we would reflect that dependency in Buildroot as well. It is 
really a runtime dependency, not a build-time dependency, so I added this select 
like we do for runtime dependencies.

  Of course, I didn't notice that we don't have a target python-wheel package, 
so it doesn't actually work.

  So, dropping this select I'm pretty sure is going to break something, at some 
point. I've looked a bit closer now, and it turns out that wheel is actually 
only imported in one specific function, in the "editable_wheel" command. So it's 
quite unlikely anyone will ever see this breakage.

  However, now I look a bit more in detail at it, it looks like this commit is 
in fact bogus, and Yann's original comment that the wheel dependency should be 
added to the infrastructure is in fact correct. From the documentation:

    Historically this documentation has unnecessarily listed ``wheel``
    in the ``requires`` list, and many projects still do that. This is
    not recommended. The backend automatically adds ``wheel`` dependency
    when it is required, and listing it explicitly causes it to be
    unnecessarily required for source distribution builds.

And it's pretty clear what is adding this dependency: the pep517 infra calls the 
build with

    python -m build -n -w

The "-w" stands for "build a wheel" and I guess it's obvious that that implies a 
dependency on host-python-wheel? And I guess that the relatively few pep517 
packages we have already have a setuptools-based package (and therefore 
host-python-setuptools and host-python-wheel) somewhere in their dependency chain.

  So, I do think this dependency should move the the pkg-python infra _for 
pep517 packages_, as Yann originally commented.

  Regards,
  Arnout