Indeed I had thought about that an completely forgotten about it. 0.21.0.post0 is probably enough just for the tarball package. No need to change the wheels.

From https://packaging.python.org/guides/distributing-packages-using-setuptools/,

Support for this feature is relatively recent. Your project’s source distributions and wheels (see Packaging your project) must be built using at least version 24.2.0 of setuptools in order for the python_requires argument to be recognized and the appropriate metadata generated.

In addition, only versions 9.0.0 and higher of pip recognize the python_requires metadata. Users with earlier versions of pip will be able to download & install projects on any Python version regardless of the projects’ python_requires values.

I think it should be fine. Another this is I'm not sure we should actually fail explicitly in setup.py on python2 or not.

I think this should fix it, but it would be good if someone could double check.

0.21.0.post0 is probably enough just for the tarball package.

Right, if one can remove just one file from PyPy without removing the release that could work.

We can check locally by:

• changing the version number to 0.21.0.post0 (+ your fix)
• building the sdist
• create a new empty venv for python 2.7
• pip install -U -f /path/to/dist/folder scikit-learn

It should ignore the local tarball and install the latest python 2.7 compatible wheel (scikit-learn 0.20.3) from PyPI.

The same command from an empty python 3 venv should install from the local sdist.

Hum, that might not be enough:

(py2tmp) ogrisel@ici:~/code/scikit-learn/dist$ pip install -U -f . scikit-learn
DEPRECATION: Python 2.7 will reach the end of its life on January 1st, 2020. Please upgrade your Python as Python 2.7 won't be maintained after that date. A future version of pip will drop support for Python 2.7.
Looking in links: .
Collecting scikit-learn
    ERROR: Complete output from command python setup.py egg_info:
    ERROR: Traceback (most recent call last):
      File "<string>", line 1, in <module>
      File "/tmp/pip-install-1B7AMh/scikit-learn/setup.py", line 30, in <module>
        % (platform.python_version(), sys.executable))
    RuntimeError: Scikit-learn requires Python 3.5 or later. The current Python version is 2.7.16 installed in /home/ogrisel/miniconda3/envs/py2tmp/bin/python.
    ----------------------------------------
ERROR: Command "python setup.py egg_info" failed with error code 1 in /tmp/pip-install-1B7AMh/scikit-learn/

Actually my test is probably not doing what it's supposed to do and is probably not going to work as long as the broken tarball is the most recent tarball uploaded on pypi.org ...

We might need to delete the 0.21.0 source dist from pypi.org

We might need to delete the 0.21.0 source dist from pypi.org

I also think so #13860 (comment)

I will push a fix to your PR soon (to remove the numpy / python 3 dependency just to get egg_info).

If you try to pip install a tarball generated with the setup.py from this PR in a python 2.7 environment, you will get the expected error message:

ERROR: scikit-learn requires Python '>=3.5' but the running Python is 2.7.16

generated by pip's own check_dist_requires_python function which seems correct to me.

To get more details in your tests you can do:

cd dist/
pip install -v --no-cache -U -f file://$PWD scikit-learn

Unfortunately it does not seem to fallback to the online PyPI index but it's probably expected when using -f on a local folder. So I am not sure how to test the full things without making a new sdist release (0.21.0.post0) on PyPI (and deleting the 0.21.0 tarball from PyPI).

I just deleted the 0.20.0 tarball from PyPI to avoid causing this problem to python 2 users and CI servers around the world while we are fixing the metadata.

A better long term solution that does not involve keeping Python 2 compat for the python setup.py egg_info command might be to include a pyproject.toml file in our tarball with all the necessary metadata.

I am not sure how to do that without introducing too much redundancy so better keep that for the next release.