CI Remove Windows 32 bit support

[thomasjpfan]

Reference Issues/PRs

Addresses #24612 for Windows 32 bit

What does this implement/fix? Explain your changes.

Using PIP_PREFER_BINARY=1 configures pip to prefer older binary packages over building from source.

Any other comments?

This PR only fixes the Windows 32 bit issue and not the Windows 64 bit issue.

[cmarmo]

Thanks @thomasjpfan .... I guess this is the answer to my question (see comment)

Please note that the amd64 Windows issue is fixed by adding a new dll as in https://github.com/scikit-learn/scikit-learn/pull/24446/files#diff-e0a07d85249fa03c06895043f2d57076b4a1a30b195631e1682b5470db6a5ac3 : should I offer a separate PR for that?

[thomasjpfan]

Please note that the amd64 Windows issue is fixed by adding a new dll as in #24446 (files) : should I offer a separate PR for that?

Yea, I like the idea of opening a separate PR to fix the current Windows 64bit wheels. It should also make the Python 3.11 PR smaller.

[ogrisel]

LGTM. Thanks for the fix.

[ogrisel]

Although in retrospect, won't this hide problems faced by our users that use the pip commandline with the default configuration?

Still getting a more stable wheel CI is nice. So still +0 for merging this.

[cmarmo]

Although in retrospect, won't this hide problems faced by our users that use the pip commandline with the default configuration?

This is an issue when producing scikit-learn wheels for windows 32.
The cleanest way to fix it is to specify dependencies in the pyproject.toml, but I couldn't find a way to specify the bitness of the system.
Let merge this and let windows 32bit wheels slowly die as scipy is doing.

[thomasjpfan]

Although in retrospect, won't this hide problems faced by our users that use the pip commandline with the default configuration?

Here are some options:

1. Document the work around it
2. Introduce an upper bound for SciPy on Windows 32bit
3. Wait for --only-binary be the default: Speculative: --only-binary by default? pypa/pip#9140
4. Drop Windows 32bit. Since SciPy 1.9.2 no ships Windows 32bit wheels, then we should not either.

I can get behind Option 4 and dropping Windows 32bit.

[cmarmo]

I can get behind Option 4 and dropping Windows 32bit.

+1

[thomasjpfan]

I updated this PR to remove Windows 32 bit support. If SciPy is not building for it anymore, then I think we can follow them.

[jjerphan]

There are Windows 32-specific adaptation left in some part of the code-base.

For instance, there are references and specific setup:

• in conda lock-files' fixtures:
  

  scikit-learn/build_tools/update_environments_and_lock_files.py

  Lines 348 to 354 in 5ceb8a6

  	# The Windows 32bit build use 3.8.10. No cross-compilation support for
  	# pip-compile, we are going to assume the pip lock file on a Linux
  	# 64bit machine gives appropriate versions
  	"python_version": "3.8.10",
  	"package_constraints": {
  	"scipy": "1.9.1", # 1.9.2 not available for 32 bit Windows
  	},

• in Windows wheels testing:
  

  scikit-learn/build_tools/github/test_windows_wheels.sh

  Lines 9 to 14 in fe62250

  	if [[ "$BITNESS" == "32" ]]; then
  	# 32-bit architectures use the regular
  	# test command (outside of the minimal Docker container)
  	cp $CONFTEST_PATH $CONFTEST_NAME
  	python -c "import sklearn; sklearn.show_versions()"
  	pytest --pyargs sklearn

• for the 32bit OpenBLAS which seems to use Windows (inferred from pip-windows)?
  

  scikit-learn/azure-pipelines.yml

  Lines 283 to 288 in 0ae132f

  	py38_pip_openblas_32bit:
  	DISTRIB: 'pip-windows'
  	PYTHON_VERSION: '3.8'
  	PYTHON_ARCH: '32'
  	LOCK_FILE: ./build_tools/azure/py38_pip_openblas_32bit_lock.txt
  	SKLEARN_TESTS_GLOBAL_RANDOM_SEED: '8' # non-default seed

  
  

  scikit-learn/build_tools/azure/test_script.sh

  Lines 12 to 13 in 0ae132f

  	elif [[ "$DISTRIB" == "pip-windows" ]]; then
  	source $VIRTUALENV/Scripts/activate

• for compilers flags for handling OpenMP:
  

  scikit-learn/sklearn/_build_utils/openmp_helpers.py

  Lines 24 to 27 in 594b1f7

  	if sys.platform == "win32" and ("icc" in compiler or "icl" in compiler):
  	return ["/Qopenmp"]
  	elif sys.platform == "win32":
  	return ["/openmp"]

• for this test adaptation:
  

  scikit-learn/sklearn/decomposition/tests/test_sparse_pca.py

  Lines 142 to 153 in 2c2ee2d

  	if sys.platform == "win32": # fake parallelism for win32
  	import joblib
  	_mp = joblib.parallel.multiprocessing
  	joblib.parallel.multiprocessing = None
  	try:
  	spca = MiniBatchSparsePCA(
  	n_components=3, n_jobs=2, alpha=alpha, random_state=0
  	)
  	U2 = spca.fit(Y).transform(Y)
  	finally:
  	joblib.parallel.multiprocessing = _mp

• for a test skipping reason:
  

  scikit-learn/sklearn/conftest.py

  Lines 148 to 152 in b2632de

  	elif sys.platform.startswith("win32"):
  	reason = (
  	"doctests are not run for Windows because numpy arrays "
  	"repr is inconsistent across platforms."
  	)

Should some of them be removed?

[thomasjpfan]

I think "win32" means "Windows" and does not specify the bitness. In Python, the platform is defined here: https://github.com/python/cpython/blob/0895c2a066c64c84cab0821886dfa66efc1bdc2f/PC/pyconfig.h#L288-L292

[jjerphan]

How confusing. 🤷

Then…

[jjerphan]

LGTM!

[ogrisel]

I am fine with the 32 bit windows support as long as scipy does not support it anymore.

I think we should still have at least one 32 bit runtime in our CI build (but this is the case with linux) to make sure that our Cython code is not bitness dependent to be able to support platforms such as WASM in the future.

[ogrisel]

LGTM!