Thank you @thomasjpfan for your PR!

I would prefer to set these environment variables:

CIBW_MANYLINUX_X86_64_IMAGE: manylinux1
CIBW_MANYLINUX_I686_IMAGE: manylinux1

To build the wheels for manylinux1. I do not think that we need to build for both, manylinux1 and manylinux2010, since manylinux1 wheels works for manylinux2010.

I think the advantage of building both is that we can be sure that we are able to build the "newer" version of manylinux***. Since manylinux2010 and manylinux2014 both require pip >=19.*, we can most get away with building manylinux1 and manylinux2014. This is what scipy uploads to pypi.

I think the advantage of building both is that we can be sure that we are able to build the "newer" version of manylinux***. Since manylinux2010 and manylinux2014 both require pip >=19.*, we can most get away with building manylinux1 and manylinux2014. This is what scipy uploads to pypi.

If I am not wrong, scipy only builds manylinux2014 for the aarch64 architecture and manylinux1 otherwise.

Edit: numpy builds wheels for manylinux1 and manylinux2010, so I am okay with building both. But would not be easier to set:

manylinux_image: [manylinux1, manylinux2010]

in the already defined build matrix and use:

exclude:
   - os: windows-latest
     manylinux_image: manylinux_1
   - os: windows-latest
     manylinux_image: manylinux_2010
   - os: macos-latest
     manylinux_image: manylinux_1
   - os: macos-latest
     manylinux_image: manylinux_2010

By this way, we would not need to duplicate code.

Off topic: I am missing the aarch64 wheels for scikit-learn==0.24.1. Are we out of travis credits?

If I am not wrong, scipy only builds manylinux2014 for the aarch64 architecture and manylinux1 otherwise.

Ah you are correct, numpy builds the full matrix

Off topic: I am missing the aarch64 wheels for scikit-learn==0.24.1. Are we out of travis credits?

Yea we are out of credits :(

If I am not wrong, scipy only builds manylinux2014 for the aarch64 architecture and manylinux1 otherwise.

Ah you are correct, numpy builds the full matrix

I edit my original answer with a possible idea to avoid code duplication.

Off topic: I am missing the aarch64 wheels for scikit-learn==0.24.1. Are we out of travis credits?

Yea we are out of credits :(

😢

Before we merge this, I would want to make sure this actually resolves the original issue. In the original issue, the readthedocs instance is using pip 20.3, which, in principle, should have fetched manylinux2010.

We can upload the file in test.pypi.org and have an environment to try to fetch it.
However, we need a system similar to the one in readthedoc (because for me it already fetches the wheel without issue)

FYI in the original issue, I just think that the build was starting at the moment of the upload and only the tar.gz file was available. The current builds are fetching the manylinux2010 wheels. However, we need to have pip>=19.3 that might not be always the case.

I tagged this PR for 0.24.1 as we can backport this to the 0.24.X branch and trigger a build there to generate manylinux1 artifacts and upload them manually with twine.

I will manage the backport.