On 11/10/21 2:22 pm, Ralf Gommers wrote:

On Mon, Oct 11, 2021 at 9:52 AM Laurie O <laurie_opperman@hotmail.com <mailto:laurie_opperman@hotmail.com>> wrote:

CI request: build PEP 656 (https://www.python.org/dev/peps/pep-0656/ <https://www.python.org/dev/peps/pep-0656/>) compliant "musllinux" wheels. This will allow easy installation in Alpine Docker images. GitHub issue: https://github.com/numpy/numpy/issues/20089 <https://github.com/numpy/numpy/issues/20089>

Thanks for asking Laurie. I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel. For now I'm inclined to say that https://github.com/alpine-wheels <https://github.com/alpine-wheels> has NumPy wheels you can use, and that Alpine is a niche platform, so this is quite low-prio for us.

In general, providing wheels for every platform under the sun is not sustainable - see https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H... <https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H...> for more context.

Cheers, Ralf

I am a little worried about releasing wheels untested that will then come back to use as issues with Nans, error states, or inaccurate results. We should have some verification that the platform passes the test suite, preferably a specific run in the main CI. Matti

Hi, It will probably soon be trivial to do a manylinux Alpine build / test on multibuild. There's a work on progress PR here, https://github.com/matthew-brett/multibuild/pull/430 that only requires someone to build a trivial test container, in order to get merged. It's also generally very easy to test locally with a minimal Docker setup. Cheers, Matthew On Mon, Oct 11, 2021 at 2:01 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:

On Mon, Oct 11, 2021 at 2:10 PM Matti Picus <matti.picus@gmail.com> wrote:

...

On 11/10/21 2:22 pm, Ralf Gommers wrote:

...

On Mon, Oct 11, 2021 at 9:52 AM Laurie O <laurie_opperman@hotmail.com <mailto:laurie_opperman@hotmail.com>> wrote:

CI request: build PEP 656 (https://www.python.org/dev/peps/pep-0656/ <https://www.python.org/dev/peps/pep-0656/>) compliant "musllinux" wheels. This will allow easy installation in Alpine Docker images. GitHub issue: https://github.com/numpy/numpy/issues/20089 <https://github.com/numpy/numpy/issues/20089>

Thanks for asking Laurie. I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel. For now I'm inclined to say that https://github.com/alpine-wheels <https://github.com/alpine-wheels> has NumPy wheels you can use, and that Alpine is a niche platform, so this is quite low-prio for us.

In general, providing wheels for every platform under the sun is not sustainable - see https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H... <https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H...> for more context.

Cheers, Ralf

I am a little worried about releasing wheels untested that will then come back to use as issues with Nans, error states, or inaccurate results. We should have some verification that the platform passes the test suite, preferably a specific run in the main CI.

I agree. Platforms where no maintainer is able to build/test locally to resolve issues are already annoying, and not having regular CI is worse. That's why I'm not sure we want to provide Alpine wheels even if our wheel build machinery setup improves. In the end it becomes a maintainer bandwidth issue.

Cheers, Ralf

_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: matthew.brett@gmail.com

Hi, On Mon, Oct 11, 2021 at 3:05 PM Matti Picus <matti.picus@gmail.com> wrote:

On 11/10/21 4:38 pm, Matthew Brett wrote:

...

Hi,

It will probably soon be trivial to do a manylinux Alpine build / test on multibuild. There's a work on progress PR here,

https://github.com/matthew-brett/multibuild/pull/430

that only requires someone to build a trivial test container, in order to get merged.

It's also generally very easy to test locally with a minimal Docker setup.

A PR to https://github.com/multi-build/docker-images to add Apline images with a complement of python versions would be welcome. The motivation for not using the muslinux containers provided by pypa/manylinux is to check that the wheel built on the pypa/manylinux containers will work in the full distro.

My guess is that that would be less than 90 minutes of work for someone with some Docker experience. It basically involves some edits to the existing build script for Pythons on the current Intel test images. See this directory: https://github.com/multi-build/docker-images/tree/focal/docker and the PR linked above for details. Please do ask for help if stuck. Cheers, Matthew

Hi Stefan, If you haven't got to it, I can do it. Cheers, Matthew On Tue, Oct 12, 2021 at 1:13 AM Stefan van der Walt <stefanv@berkeley.edu> wrote:

On Mon, Oct 11, 2021, at 07:56, Matthew Brett wrote:

...

My guess is that that would be less than 90 minutes of work for someone with some Docker experience.

I'll give it a shot.

Stéfan _______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: matthew.brett@gmail.com

I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel

I've been really impressed with cibuildwheel. It works well on my project, so I'd be interested to see how it goes on something with a more complex build (numpy/scipy). Is there a plan for such a transition?

On Mon, Oct 11, 2021 at 2:55 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:

On Mon, Oct 11, 2021 at 2:18 PM Andrew Nelson <andyfaff@gmail.com> wrote:

...

...

I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel

I've been really impressed with cibuildwheel. It works well on my project, so I'd be interested to see how it goes on something with a more complex build (numpy/scipy). Is there a plan for such a transition?

A wish, but not a real plan just yet.

I stand corrected, it turns out Thomas Fan had exactly such a plan:) https://github.com/numpy/numpy/pull/20102 Cheers, Ralf

On 11/10/21 2:22 pm, Ralf Gommers wrote:

On Mon, Oct 11, 2021 at 9:52 AM Laurie O <laurie_opperman@hotmail.com <mailto:laurie_opperman@hotmail.com>> wrote:

CI request: build PEP 656 (https://www.python.org/dev/peps/pep-0656/ <https://www.python.org/dev/peps/pep-0656/>) compliant "musllinux" wheels. This will allow easy installation in Alpine Docker images. GitHub issue: https://github.com/numpy/numpy/issues/20089 <https://github.com/numpy/numpy/issues/20089>

Thanks for asking Laurie. I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel. For now I'm inclined to say that https://github.com/alpine-wheels <https://github.com/alpine-wheels> has NumPy wheels you can use, and that Alpine is a niche platform, so this is quite low-prio for us.

In general, providing wheels for every platform under the sun is not sustainable - see https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H... <https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H...> for more context.

Cheers, Ralf

I am a little worried about releasing wheels untested that will then come back to use as issues with Nans, error states, or inaccurate results. We should have some verification that the platform passes the test suite, preferably a specific run in the main CI. Matti

Hi, It will probably soon be trivial to do a manylinux Alpine build / test on multibuild. There's a work on progress PR here, https://github.com/matthew-brett/multibuild/pull/430 that only requires someone to build a trivial test container, in order to get merged. It's also generally very easy to test locally with a minimal Docker setup. Cheers, Matthew On Mon, Oct 11, 2021 at 2:01 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:

On Mon, Oct 11, 2021 at 2:10 PM Matti Picus <matti.picus@gmail.com> wrote:

...

On 11/10/21 2:22 pm, Ralf Gommers wrote:

...

On Mon, Oct 11, 2021 at 9:52 AM Laurie O <laurie_opperman@hotmail.com <mailto:laurie_opperman@hotmail.com>> wrote:

CI request: build PEP 656 (https://www.python.org/dev/peps/pep-0656/ <https://www.python.org/dev/peps/pep-0656/>) compliant "musllinux" wheels. This will allow easy installation in Alpine Docker images. GitHub issue: https://github.com/numpy/numpy/issues/20089 <https://github.com/numpy/numpy/issues/20089>

Thanks for asking Laurie. I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel. For now I'm inclined to say that https://github.com/alpine-wheels <https://github.com/alpine-wheels> has NumPy wheels you can use, and that Alpine is a niche platform, so this is quite low-prio for us.

In general, providing wheels for every platform under the sun is not sustainable - see https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H... <https://mail.python.org/archives/list/numpy-discussion@python.org/thread/46H...> for more context.

Cheers, Ralf

I am a little worried about releasing wheels untested that will then come back to use as issues with Nans, error states, or inaccurate results. We should have some verification that the platform passes the test suite, preferably a specific run in the main CI.

I agree. Platforms where no maintainer is able to build/test locally to resolve issues are already annoying, and not having regular CI is worse. That's why I'm not sure we want to provide Alpine wheels even if our wheel build machinery setup improves. In the end it becomes a maintainer bandwidth issue.

Cheers, Ralf

_______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: matthew.brett@gmail.com

Hi, On Mon, Oct 11, 2021 at 3:05 PM Matti Picus <matti.picus@gmail.com> wrote:

On 11/10/21 4:38 pm, Matthew Brett wrote:

...

Hi,

It will probably soon be trivial to do a manylinux Alpine build / test on multibuild. There's a work on progress PR here,

https://github.com/matthew-brett/multibuild/pull/430

that only requires someone to build a trivial test container, in order to get merged.

It's also generally very easy to test locally with a minimal Docker setup.

A PR to https://github.com/multi-build/docker-images to add Apline images with a complement of python versions would be welcome. The motivation for not using the muslinux containers provided by pypa/manylinux is to check that the wheel built on the pypa/manylinux containers will work in the full distro.

My guess is that that would be less than 90 minutes of work for someone with some Docker experience. It basically involves some edits to the existing build script for Pythons on the current Intel test images. See this directory: https://github.com/multi-build/docker-images/tree/focal/docker and the PR linked above for details. Please do ask for help if stuck. Cheers, Matthew

Hi Stefan, If you haven't got to it, I can do it. Cheers, Matthew On Tue, Oct 12, 2021 at 1:13 AM Stefan van der Walt <stefanv@berkeley.edu> wrote:

On Mon, Oct 11, 2021, at 07:56, Matthew Brett wrote:

...

My guess is that that would be less than 90 minutes of work for someone with some Docker experience.

I'll give it a shot.

Stéfan _______________________________________________ NumPy-Discussion mailing list -- numpy-discussion@python.org To unsubscribe send an email to numpy-discussion-leave@python.org https://mail.python.org/mailman3/lists/numpy-discussion.python.org/ Member address: matthew.brett@gmail.com

I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel

I've been really impressed with cibuildwheel. It works well on my project, so I'd be interested to see how it goes on something with a more complex build (numpy/scipy). Is there a plan for such a transition?

On Mon, Oct 11, 2021 at 2:55 PM Ralf Gommers <ralf.gommers@gmail.com> wrote:

On Mon, Oct 11, 2021 at 2:18 PM Andrew Nelson <andyfaff@gmail.com> wrote:

...

...

I'd say we may perhaps want to do this, but only if and when it becomes much easier - for example once we have migrated all our wheel build infra to GitHub Actions and cibuildwheel

I've been really impressed with cibuildwheel. It works well on my project, so I'd be interested to see how it goes on something with a more complex build (numpy/scipy). Is there a plan for such a transition?

A wish, but not a real plan just yet.

I stand corrected, it turns out Thomas Fan had exactly such a plan:) https://github.com/numpy/numpy/pull/20102 Cheers, Ralf