It would be nice to upload the Pyodide wheel to anaconda.org when the tests pass, like numpy does. I guess this can be done in a separate PR.

The middle-term goal would be to use the Pyodide wheel on anaconda.org inside our doc JupyterLite.

Yes, I left it for a follow-up PR, indeed. On being able to use the Pyodide wheel, there are two approaches I have in mind:

• use the built wheel from Anaconda.org (pre-download it using pip or cURL) for the nightly wheels (requires Add a --index-urls CLI flag for pip-install magics jupyterlite/pyodide-kernel#166)
• Build the wheel separately again (which doesn't take long – it is the tests that take much longer) as a part of the docs build, leave it for JupyterLite to find, and modify the notebook modification function to install that wheel for all the notebooks. This means that the notebooks in the docs all versions going forward can have a scikit-learn wheel to install which will be synced with the docs version. This is the approach I've taken in Add JupyterLite-powered interactive galleries to the scikit-image documentation scikit-image/scikit-image#7644 that you reviewed some time back ;) (I'd be grateful for another look there!)
• or, a combination of both in the second approach: don't build the wheel separately but download it from Anaconda.org, leave it in for JupyterLite to find, and follow the same notebook modification functions

Which one of these would you prefer?

My preference would be the following for now because I feel this is the simplest option (we can always change our mind later):

• upload the Pyodide wheel to anaconda.org after this worflow succeeds (similar to what numpy I think)
• in the JupyterLite notebook, one of the first cell should do something like this (through sphinx-gallery notebook_modification_function)

  import micropip
  # this line is needed as workaround see https://github.com/pyodide/micropip/issues/223
  await micropip.install(["joblib", "threadpoolctl", "scipy"])
  await micropip.install("scikit-learn", index_urls="https://pypi.anaconda.org/scientific-python-nightly-wheels/simple", pre=True)

One of my concern for your approach in scikit-image is the size of the wheel (~12M for scikit-learn Pyodide wheel). On each commit of the repo, we will push an updated wheel and the size of the scikit-learn.github.io repo grows quite quickly. To be honest, maybe the size of the Pyodide wheel is actually quite small compared to the size of the website (~300MB) ...

This is the approach I've taken in Add JupyterLite-powered interactive galleries to the scikit-image documentation scikit-image/scikit-image#7644 that you reviewed some time back ;) (I'd be grateful for another look there!)

I'll try to have a closer look although I can't make any promises 😅

Ah, I see – the increased size of the docs repository would make it quite big (I thought new builds are being overwritten on the gh-pages branch). Also, I see that the size of the zipped docs is monotonically increasing with each version for recent versions: https://scikit-learn.org/dev/versions.html, so while it it's okay for archives, it's indeed nicer if a GitHub repository does not include any more binary files than it needs to.

I'm happy to implement the notebook modifications you suggest once this PR is merged!

Here, cache was a spurious argument that was unused throughout the file. It broke the Pyodide tests and returned an error upon collection because the interpreter assumed that this was a fixture.

This is weird __import__ is not supported in Pyodide, I just tried __import__('joblib') in Pyodide console and this fails.

What I am guessing happens is that this test somehow works because pkg.util_walkpackages populates sys.modules and __import__("bla") works inside Pyodide if "bla" is already in sys.modules. Oh well 🤷 ...

🤔 I'm not sure why that's the case. I tried it just now – both in JupyterLite and in the Pyodide (stable) console and I was able to __import__("joblib") to get <module 'joblib' from '/lib/python3.12/site-packag es/joblib/__init__.py'>. I did need to install joblib with micropip, of course. Could you please check this again, just in case I might be missing something?

Either way, it's nice that the test now works – I hope!

I think micropip.install('joblib') makes sys.modules['joblib'] exist and so __import__ work.

In any case this is not really important since the test appears to pass now 🤷.