Add Support for Message Based Interrupt

[afshin]

Implements handling of interrupt_request message, which is now possible due to the control channel being on a separate thread.
Does not implement handling on Windows, which is less likely to be used for remote kernels.

[blink1073]

Excellent, thank you!

[blink1073]

cc @SylvainCorlay @kevin-bates

[SylvainCorlay]

Thanks for this!
Happy to see ipykernel move closer to fully implementing the protocol ;)

[ccordoba12]

I think this is not enough because it's missing a call to perform the request in jupyter_client, just like there's one for shutdown:

https://github.com/jupyter/jupyter_client/blob/9c0b4c0e28a9c17d9c14123522415b2048bd9522/jupyter_client/client.py#L762-L781

Otherwise, how is it possible to perform an interrupt from the kernel client?

[kevin-bates]

Hi @ccordoba12 - the interrupt machinery comes from the KernelManager by virtue of the kernelspec indicating support for message-based interrupts.

Hmm - should we make signal-based interrupts the else so that kernelspecs that don't specify a mode attempt at least that form of interrupt?

[blink1073]

According to the docs, it is opt-in: "For this to work, the kernels kernelspec must set interrupt_mode to message". I think the question is whether the ipykernel default spec should be message based.

[kevin-bates]

ok - I understood the question to be one of missing logic - sorry about that.

My question was more general and I just found the answer in that 'signal' will be used if interrupt_mode is not specified, which, now that I recall, have seen before.

I'll let the maintainers decide what the default mode should be. 😄

[ccordoba12]

the interrupt machinery comes from the KernelManager by virtue of the kernelspec indicating support for message-based interrupts.

Yeah, but if you connect to a kernel started elsewhere (for example, locally to a kernel in a server, which Spyder users do a lot), then you don't have access to a kernel manager, just a kernel client. And in that case it'd be really nice to be able to send an interrupt message using the control channel (like you can now to perform a shutdown with the method I referenced above).

I thought this PR was about that, but it seems I was mistaken, sorry.

[kevin-bates]

I'm sorry, I probably steered this discussion in the wrong direction and I apologize. Please continue.

[ccordoba12]

No worries @kevin-bates. Let's ping @davidbrochart to see what he thinks about this one.

@davidbrochart, do you think it's worth it to add a message-based interrupt in jupyter-client?

[SylvainCorlay]

I think it is worth it just for the sake of completeness.

[ccordoba12]

Agreed. I'll open a PR in jupyter-client so that (hopefully) it can be part of its 7.0 version.