Pexpect-like interfaces based on jupyter_client

[mkoeppe]

Many computational systems are nowadays available as jupyter kernels.
We should interface to them using jupyter_client.
This would be more robust than pexpect.

Related:

• Alert pexpect interfaces to library interfaces #22680: Alert pexpect interfaces to library interfaces
• Meta-ticket: replace pexpect interface calls with library where possible #16688: Replace pexpect interface calls with library where possible

CC: @jdemeyer @simon-king-jena @sebasguts

Component: interfaces

Keywords: days84

Issue created by migration from https://trac.sagemath.org/ticket/22579

[jdemeyer]

comment:1

Do you have a concrete kernel in mind? I'm wondering if there are any kernels where this approach makes more sense than other approaches (like pexpect or directly interfacing the libary).

[mkoeppe]

comment:2

This topic came up here at days84 related to polymake (#22452), however I see very wide applicability.

Obviously, directly interfacing a library is almost always better -- except if one wants to be able to access a remote server.

But a pexpect interface seems to be much easier to write -- except for tricky/brittle things such as tab completion and signal handling.

With a jupyter kernel all these tricky things are already implemented. So there would be a separation of concerns, which we can benefit from.

Moreover, since there is a clear over-the-wire protocol, rather than having to send everything through a pty, this will be more robust.

[jdemeyer]

comment:3

I think it all depends on how the Jupyter kernel is implemented. Obviously, if the Jupyter kernel uses pexpect, you cannot gain much from this approach.

[jdemeyer]

comment:6

If nobody can come up with a concrete proposal here, I'd rather close this as "wontfix". It's a cute idea but I don't really see the point.

[mkoeppe]

comment:7

I have set it to 'wishlist'. I hope that's good enough. I would like to keep the ticket open because I may want to work on it at some time.