Use `coincident` if `crossOriginIsolated`, `comlink` otherwise

[jtpio]

Looking into https://github.com/jupyterlite/jupyterlite/issues/1424.

coincident uses a SharedArrayBuffer here: https://github.com/WebReflection/coincident/blob/01cc8a1e90abb360deb9eabd7773ef4355fcbc3d/esm/index.js#L82

However in environments that are not crossOriginIsolated, usings coincident on the worker side leads to the following issues:

TypeError: [object Int32Array] is not a shared typed array.

So this PR takes the following approach:

• use coincident on both the main thread and the worker if `crossOriginIsolated, to fix https://github.com/jupyterlite/jupyterlite/issues/1424 / #127
• use comlink otherwise, this time on both the main thread and the worker too

Fixes https://github.com/jupyterlite/pyodide-kernel/issues/127

[github-actions[bot]]

:point_left: Try it on ReadTheDocs

[jtpio]

Current status, after a naive replacement of postMessage by coincident:

[bollwyvl]

Maybe we need to be using coincident/structured?

[jtpio]

Right, I'll have a closer look.

[jtpio]

This seems to be caused by coincident using a SharedArrayBuffer here: https://github.com/WebReflection/coincident/blob/01cc8a1e90abb360deb9eabd7773ef4355fcbc3d/esm/index.js#L82

It seems to be working fine when serving the site with right headers, for example with: npx static-handler --cors --coop --coep --corp ./build/docs-app

I thought coincident would default to the comlink behavior when SharedArrayBuffer are not available, but it looks like it's not the case.

[jtpio]

Switching to coincident on the worker side does seem to be fixing https://github.com/jupyterlite/jupyterlite/issues/1424.

coincident-widget-error.webm

I thought coincident would default to the comlink behavior when SharedArrayBuffer are not available, but it looks like it's not the case.

Maybe we need to clearly separate the case when SharedArrayBuffer are available on the page from when they are not. Meaning using coincident when they are, and comlink when they are not (on both the worker and the main threads).

cc @martinRenou in case you have some input on this

[martinRenou]

Meaning using coincident when they are, and comlink when they are not (on both the worker and the main threads).

That's unfortunate, but if it's the only solution than let's go for it.

coincident is really taking the burden of implementing a proper communication protocol using SharedArrayBuffers, so it's really useful to us in that case. That's unfortunate it cannot work in an asynchronous fashion though, closer to the comlink approach.

[martinRenou]

Thanks for looking into it by the way 👍🏽 I'm swamped with other projects. We should apply similar changes to jupyterlite/xeus once we have a proper solution.

[bollwyvl]

Yeah... As realizing the merit of the SAB is getting more and more tenuous, it looks like it needs to be entirely opt-in.

On Wed, Jul 17, 2024, 03:42 martinRenou @.***> wrote:

Thanks for looking into it by the way 👍🏽 I'm swamped with other projects. We should apply similar changes to jupyterlite/xeus once we have a proper solution.

— Reply to this email directly, view it on GitHub https://github.com/jupyterlite/pyodide-kernel/pull/126#issuecomment-2232762772, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAALCRHRBWNJL2KS5OEY7I3ZMYVBBAVCNFSM6AAAAABKYZQMG6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDEMZSG43DENZXGI . You are receiving this because you commented.Message ID: @.***>

[jtpio]

Looks like we could keep the current logic of going over the SharedArrayBuffer if they are available, and default to the service worker when they are not.

If the SharedArrayBuffers are available, we use coincident on both the main thread and the worker. If they are not we use comlink.

[jtpio]

I'll push a first draft towards that direction and will clean things up a bit afterwards.

[jtpio]

Thanks for the review!

[martenrichter]

This PR caused the breakage of interact, see: https://github.com/jupyterlite/pyodide-kernel/issues/143