using voila specific kernel manager

[maartenbreddels]

Now voila and the server extension both use a voila_kernel_manager instead of kernel_manager. This means that voila used as server extension also does not allow arbitrary code execution.

[maartenbreddels]

Also allow the custom_message for voila standalone and server extension.

[jtpio]

Thinking about this a bit more, having a separate kernel manager for voila could also be very useful to:

• show the "voila only" kernels (https://github.com/voila-dashboards/voila/pull/611), as clients could send a GET /voila/api/kernels when voila is used as a server extension, and will only get the voila kernels without the regular kernel sessions started via the notebook frontend
• configure the idle culler to only cull the voila kernels when voila is used as a server extension (this was suggested by @maartenbreddels).

[jtpio]

I can rebase this PR onto the latest master tomorrow.

(cc @maartenbreddels let me know if you have any objections)

[maartenbreddels]

Yes, please go ahead, I think we should have a different manager, with different defaults, different permissions, and different authentication.

[maartenbreddels]

Something that crossed my mind: Would it be a different pool of kernels? Like would a kernel spawned by notebook or lab be visible to voila? I'm thinking about a future use case where you have a running notebook, and want to share that existing kernel via voila.

[jtpio]

I'm thinking about a future use case where you have a running notebook, and want to share that existing kernel via voila.

It looks like this should be possible at the moment with voila running as a server extension sharing the same kernel manager as the notebook server, by passing an existing kernel_id to the start_kernel?

https://github.com/voila-dashboards/voila/blob/b75e05f4041aaaf91b3ee9e3c08848136ebec5a3/voila/handler.py#L120

Although there will probably be some issues with the widgets (multiple clients for the same kernel)

Would it be a different pool of kernels? Like would a kernel spawned by notebook or lab be visible to voila?

If we isolate the voila kernel manager, then this becomes a bit trickier. Unless the voila kernel manager is able to proxy some of the messages to the other kernel manager (not sure yet if that's even possible right now).

[maartenbreddels]

Although there will probably be some issues with the widgets (multiple clients for the same kernel)

Currently yes, but there are ways around that, and for sure it's going to be supported in the future :)

[SylvainCorlay]

I don't think that this resolves the arbitrary code execution issue though, because the raw jupyter server always exposes the base kernel API endpoint.

[jtpio]

A couple of things to cover before this change can be considered as "ready":

• [ ] Fix failing test in tests/server/cwd_subdir_test.py (for reference: https://github.com/voila-dashboards/voila/runs/720801386?check_suite_focus=true)
• [ ] Shutdown all voila kernels on Ctrl-C / SIGTERM when voila is used as server extension (they are not shut down automatically anymore now that there is a separate kernel manager instance)
• [ ] Add test for the kernel namespacing when voila is used as a server extension (/api/kernels and /voila/api/kernels should return different results)

[jtpio]

Shutdown all voila kernels on Ctrl-C / SIGTERM when voila is used as server extension (they are not shut down automatically anymore now that there is a separate kernel manager instance)

The ExtensionApp.stop() method might be relevant here (although it seems to be specific to standalone mode?):

https://github.com/jupyter/jupyter_server/blob/f1df254e17854aecb0ef025a5891c6910c62fd70/jupyter_server/extension/application.py#L387-L391

[trungleduc]

Closing as the proposed changes are not applicable anymore to the next release of Voila