You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Wheels on PyPI violate manylinux1 specification / PEP513 #8802
As already pointed out by @jjhelmus, the Linux wheels on PyPI do not conform to the manylinux1 specification provided in PEP513. To quote @jjhelmus's comment from #5033:
Specifically, they contain shared libraries which reference versioned symbols which do not meet the GLIBC <= 2.5 and GLIBCXX <= 3.4.9 requirements specified in the PEP.
It would be great if the wheels could be fixed to match the specification.
@yifeif From what I understand on the issue, it was decided to just leave a note on the website?
Maybe we should just remove manylinux1 wheel files, as they are leading to confusion.
As we only officially support ubuntu officially, is there a way to mark the wheel files as such, instead of the broader manylinux label?
I don't understand the motivation to label the wheel "manylinux" combined with a note that it actually isn't a manylinux wheel. If it does not follow the specification, it's just a regular wheel and calling it manylinux is confusing.
To maximise compatibility, it would obviously be much preferred to follow the PEP and build the wheels against the official manylinux docker images (unless there is a true incompatibility).
@bluenote10 thanks for raising the concern. We are not building against the official manylinux docker for many reasons such as we need nvidia docker to build GPU binaries, we don't have support for centOS etc. And manylinux1 is the only linux ABI Pypi supports now. We'd like to provide the convenience of downloading linux wheel through pip install for now until better solution comes up.
Just chiming in here. It's counter-productive to claim your wheels are manylinux1-compliant when they aren't. This is producing annoyances for other projects in the ecosystem who have to deal with bug reports such as "pyarrow crashes when I try to use it alongside tensorflow".
Reacted by dave-shawley, Pradyun Gedam, Tony Kelman, Xavier Fernandez, Juan Luis Cano Rodríguez, Keith Kraus, Joachim Folz, Stella Laurenzo and Peter Výboch
The problem I have with is if I build and publish my own wheel internally, and use an extra-index-url to my pypi repo, the public pypi 'manylinux1' will be selected by pip over my internal 'linux' wheel. If the published wheels were manylinux1, I wouldnt need to build my own.
Activity
gunan commentedon Mar 29, 2017
@yifeif From what I understand on the issue, it was decided to just leave a note on the website?
Maybe we should just remove manylinux1 wheel files, as they are leading to confusion.
As we only officially support ubuntu officially, is there a way to mark the wheel files as such, instead of the broader manylinux label?
bluenote10 commentedon Apr 5, 2017
I don't understand the motivation to label the wheel "manylinux" combined with a note that it actually isn't a manylinux wheel. If it does not follow the specification, it's just a regular wheel and calling it manylinux is confusing.
To maximise compatibility, it would obviously be much preferred to follow the PEP and build the wheels against the official manylinux docker images (unless there is a true incompatibility).
yifeif commentedon Apr 5, 2017
@bluenote10 thanks for raising the concern. We are not building against the official manylinux docker for many reasons such as we need nvidia docker to build GPU binaries, we don't have support for centOS etc. And manylinux1 is the only linux ABI Pypi supports now. We'd like to provide the convenience of downloading linux wheel through pip install for now until better solution comes up.
pitrou commentedon Jul 2, 2018
Just chiming in here. It's counter-productive to claim your wheels are manylinux1-compliant when they aren't. This is producing annoyances for other projects in the ecosystem who have to deal with bug reports such as "pyarrow crashes when I try to use it alongside tensorflow".
westurner commentedon Sep 18, 2018
"PEP 571 -- The manylinux2010 Platform Tag"
https://www.python.org/dev/peps/pep-0571/
"Tracking issue for manylinux2010 rollout"
pypa/manylinux#179
"[Distutils] Opinions on requiring younger glibc in manylinux1 wheel?"
https://mail.python.org/mm3/archives/list/[email protected]/message/M4MSVY5MPAPXFWHH4PBLE6PEBPOBIA44/
chargrnv commentedon Nov 28, 2018
The problem I have with is if I build and publish my own wheel internally, and use an extra-index-url to my pypi repo, the public pypi 'manylinux1' will be selected by pip over my internal 'linux' wheel. If the published wheels were manylinux1, I wouldnt need to build my own.
@yifeif can this be reopened please?