Use jupyterlite instead of mybinder.org with the same result

[choldgraf]

I noticed that the matplotlib Binder link (mpl-brochure-binder) currently takes up about 17% of mybinder.org's traffic (about 2.1K launches last week for example). Could I ask that y'all explore using jupyterlite for this instead of Binder? I think the demos should be simple enough that the jupyterlite environment can handle it.

@jtpio any chance you could give tips on how to accomplish this? This could offload a lot of traffic from Binder to something a lot less compute-intensive.

[jtpio]

@choldgraf quick looking at the notebook I would expect it to work fine in JupyterLite.

Actually it's possible to try it with https://jupyter.org/try-jupyter/lab/?fromURL=https://raw.githubusercontent.com/matplotlib/mpl-brochure-binder/refs/heads/main/MatplotlibExample.ipynb.

And adding %pip install ipympl in a cell before.

Although there seems to be an issue coming from ipympl when executing the cell.

[timhoffm]

I get

File /lib/python3.13/site-packages/matplotlib/backends/backend_webagg_core.py:20
     17 import os
     18 from pathlib import Path
---> 20 from js import alert, document
     21 import numpy as np
     22 from PIL import Image

ImportError: cannot import name 'alert' from 'js' (unknown location)

Which I don't understand because that is our module, but the import is not in our sources https://github.com/matplotlib/matplotlib/blob/dc5615303d5eafcbae101d47fbbfa81f84172361/lib/matplotlib/backends/backend_webagg_core.py#L16-L20

@jtpio I'm not into the pyodide world. Is this some specially adapted version of matplotlib?

[story645]

I think the support for adding jupyterlite to our landing page (matplotlib/matplotlib#29506) cross applies here, especially cause of the burden we're putting on you folks (sorry!). Pretty sure either/both @ksunden and @QuLogic have worked on this at some point (xref matplotlib/matplotlib#29093)

[timhoffm]

@choldgraf since you get the traffic, do you have data on whether people are actually looking into the notebook? Something like how long they stay on the page or whether / how many cells are executed. That'd inform us how important a live demo feature is. If most people start binder but then don't actually use it, we could remove the link even if no jupyterlite replacement is ready yet.

[ianthomas23]

The webagg backend is patched in pyodide to give it a working interactive backend in standalone (outside of JupyterLite) deployments. This won't work in JupyterLite and will evidently break ipympl as well.

You should use the xeus-python kernel not the pyodide one. The former is up to date with Matplotlib 3.10.6 and works with ipympl whereas the latter is currently stuck at Matplotib 3.8.4.

[jtpio]

Opened matplotlib/mpl-brochure-binder#14

This appears to be working fine here when deployed from the main branch of my fork: https://jtpio.github.io/mpl-brochure-binder/lab/index.html?path=MatplotlibExample.ipynb

Probably it wouldn't hurt to have this merged since it would be deployed on GitHub Pages on the side.

[ksunden]

I would be interested in seeing Jupyterlite included. I do think the primary hang-up has been that none of us have dug deep enough into the different kernel options and many of the most popular options have quite outdated Matplotlib. (And we have been against putting outdated mpl on our own website).

If there is a good way of deploying xeus based kernel with up to date mpl, I'd be interested in seeing it happen.

I also know that in my (non-mpl directly) trials with Jupyterlite, there are interesting differences in how dependencies get installed (I do not recall the specifics, but believe one it was difficult to specify ahead of time, but allowed dynamic install and the other dynamic install was hard/impossible, but could be specified ahead of time)