Always install pre-built wheels

[kdmccormick]

Always install pre-built wheels

[kdmccormick]

Make sure that all of the Open edX packages get all of the appropriate wheels pushed to PyPI with each release. This will generally be a single universal wheel unless the package contains native code extensions.

For ~95% of Open edX repos, would this be as simple as setting {'bdist_wheel':{'universal':'1'}}?

If so, would that on its own be worth doing, whether or not we go for the rest of the issue?

Create a repository and Docker image(s) with all of the dependencies needed to build the missing dependency wheels. This could also be used in development to try out new packages with missing wheel variants on PyPI, building them here and then supplying them to the dev environment's devpi instance.

I think I would want to look at a list packages that are missing dependency wheels so we can get an idea of how much build time & image size we'd save. If it'd be significant, then I'm in favor of this.

Some of the benefits could be obtained by instead/also using a builder pattern for Dockerfiles

FWIW, Tutor and most of its plugins use the builder pattern, yet I still want to bring the build time and image size down.

[kdmccormick]

@jmbowman :

I think we'd want to start with a repo health check that checks PyPI for wheel availability, that would probably go a long way towards answering your questions. I suspect it's only around 2% of our package dependencies that have any native code extensions, which is why it feels silly to me that we're bloating our build process because of those.

[jmbowman]

For ~95% of Open edX repos, would this be as simple as setting {'bdist_wheel':{'universal':'1'}}?

We should already be doing this for most of them, we just need to be consistent about it. The need is for both configuration like https://github.com/openedx/django-user-tasks/blob/master/setup.cfg#L1-L2 and working automated release uploads like https://github.com/openedx/django-user-tasks/blob/master/.github/workflows/pypi-publish.yml . (And actually doing releases instead of linking to commits on GitHub.)

If so, would that on its own be worth doing, whether or not we go for the rest of the issue?

Yes, every package made available as a wheel that wasn't before should speed things up just a little bit more. A lot more in the case of the handful of (mostly upstream) packages with nontrivial native code that needs to be compiled.

I think we'd want to start with a repo health check that checks PyPI for wheel availability, that would probably go a long way towards answering your questions. I suspect it's only around 2% of our package dependencies that have any native code extensions, which is why it feels silly to me that we're bloating our build process because of those.

A couple of repo health checks could definitely help us be more consistent about this, especially if combined with edx/edx-arch-experiments#231 . And the main thing that's been forcing us to keep those build dependencies has been the lack of an official process for building those 3rd-party wheels and a public place to store them. I wanted to set that up for a long time, but it just never rose to the top of 2U's priority list.

Another reason to do this and extend it even to upstream packages that do have wheels on PyPI is protection against the sudden removal of dependencies from PyPI. This actually bit us a few times, and usually resulted in at least a half-day fire drill where a few people had to drop everything, fork the repo, get the fork released on PyPI, and untangle enough dependencies to make everything depend on the fork's name instead. (And deployments are impossible until all that gets sorted out.) That problem goes away if we maintain a package server that nobody else has permission to delete releases from.