Reducing default set of packages

[asmeurer]

I noticed that matplotlib and numpy are always loaded by default, if I understand the console output correctly.

Can we make it so that they are not loaded by default? It seems that most packages work like this, where if you import something like scipy it loads it. That would presumably make the initial loading time faster.

[jtpio]

I think this is likely fixed in the latest versions of the jupyterlite-core and jupyterlite-pyodide-kernel packages.

The relevant PR being jupyterlite/jupyterlite#911.

[asmeurer]

Well there's also the converse issue, where dependencies are not loaded unless they are imported explicitly by the user #18. Do the recent changes affect that?

[jtpio]

unless they are imported explicitly by the user #18. Do the recent changes affect that?

This is likely still the case with the latest versions. iirc Pyodide detects import statements and loads the appropriate packages accordingly.

Not sure we can have both. As an alternative we could also try using the xeus python kernel to only include a subset of packages and not rely on the default versions provided by Pyodide: https://github.com/jupyterlite/xeus-python-kernel. But with that kernel the packages will be loaded at startup as well and not lazily (at least not for now).

[asmeurer]

But why doesn't it operate on import statements from within sympy? Internally sympy.plot calls import matplotlib to see if it's installed, but that doesn't trigger installation. Only import matplotlib in the jupyterlite REPL itself does.

[jtpio]

If I remember correctly Pyodide parses the source submitted by a user and looks for import statements. But I don't think it does this for imports within a module.