Enhance the platform tag definition for wheel files to cover platform variants

[qwcode]

The current definition of the "platform" tag in PEP 425 is inadequate:

• it doesn't account for Linux distribution variations
• it doesn't clearly indicate the Windows MSVC runtime (this is technically implicit in the CPython runtime version, but could be clearer)
• it doesn't account for variations in SSE support (None, SSE2, SSE3), which is critical for NumPy

Since distutils.util.get_platform() isn't providing sufficient platform discrimination, we need to come up with a new definition that provides the details we need, while also allowing pip to automatically fall back to the less specific wheels if the exact variant isn't available.

---

migrated from https://bitbucket.org/pypa/pypi-metadata-formats/issues/15/enhance-the-platform-tag-definition-for

[soltanmm]

What's the state of this?

[wil93]

I'm also interested in the state of this issue, especially the first of the three points (accounting for Linux distribution variations). I believe it's crucial to add basic Linux support, leaving the more specific discrimination (SSE support and whatnot) for the next version.

One reason to add basic Linux support is this: when building a docker container, pip install gevent (or similar) will take a lot of time, everytime, because every pip install actually behaves like pip install --no-cache-dir. This can easily become painful: if at the end of a docker build there's an error and you have to modify the Dockerfile, the next build will again need to recompile everything (as opposed to just re-download everything).

[ncoghlan]

The "manylinux1" concept @njsmith recently posted about on distutils-sig seems like the most promising prospect for near term progress: https://mail.python.org/pipermail/distutils-sig/2016-January/027980.html

[piotr-dobrogost]

I know @ncoghlan was working on this with regard to identification of Linux distributions by means of lsb_realease and such. Nick, could you please link here relevant discussions that were taking place?

[ncoghlan]

"working on" is a generous description - "kicking ideas around" was as far I got, although Nate Coraor took things further for use in https://galaxyproject.org/

The original incarnation of the idea was this proposal to use /etc/os-release: https://mail.python.org/pipermail/distutils-sig/2014-November/025297.html

Nate then set about turning that idea into reality: https://mail.python.org/pipermail/distutils-sig/2015-July/026616.html

[StephanErb]

FWIW there is also the manylinux proposal: https://code.activestate.com/lists/python-distutils-sig/27556/ and https://www.python.org/dev/peps/pep-0513/

[ncoghlan]

Yes, we see the two proposals as complementary - manylinux1 for "one size fits most" wheels, and the more specific ones for when you want to build wheel files that are optimised for a specific distribution, but don't want them accidentally confusing matters for anyone using a different distribution.

[thedrow]

SSE/AVX availability is important for other packages as well in case they wrap C code that uses it. I wondered what could be done about it. We could just parse /etc/cpuinfo where it's available.