Improve handling of "rapid fire" completion requests

[jameskoch]

The nteract front end is far more aggressive than Jupyter notebook about requesting completions from the kernel as the user types. There's some great UX that can be unlocked if this could work more seamlessly w/ Almond.

Ideally, when Almond receives a new completion request it would quickly interrupt an in-flight one to prioritize the new/latest one.

Would need to think about how this works in the "multiple front-ends attached to a single kernel" model (I don't know much about that world).

Discussed a bit in this nteract issue.

[jameskoch]

nsc Global docs here describe that certain requests may be prioritized higher than existing requests. Hard to tell under which exact conditions this would occur, and whether, once newer requests finish, older ones would be resumed. I might try to monkey-patch my way around Ammonite's CompilerLifecycleManager.complete being synchronized and just feed concurrent requests into its Pressy and see what happens.

[rgbkrk]

"multiple front-ends attached to a single kernel"

We made the architecture for this in Jupyter and while it's handy for debugging, it's not widely used (from a ZeroMQ standpoint). However, each client should have a unique session ID so it could be keyed on that.

[alexarchambault]

@jameskoch The scala presentation compiler, which handles completions, should support concurrent completion requests, yes. Ammonite doesn't use that and just blocks until a completion result arrives.

I guess the ammonite.interp.Pressy API can be tweaked, so that we can submit concurrent requests.

I'm going to tweak the almond.interpreter.Interpreter API here so that it allows to run completions in a non-blocking and cancellable way.

[alexarchambault]

#242 allows Interpreter implementations to handle completion requests asynchronously (by overriding Interpreter.asyncComplete rather than Interpreter.complete) and make them cancellable, and automatically cancels the currently running one when a new completion request arrives.

[alexarchambault]

@jameskoch I can look into having Ammonite accept concurrent completion requests, if you haven't already.

[jameskoch]

I haven't pushed ahead on this, and I wouldn't be nearly as effective as you in doing so. Have at it!

[alexarchambault]

com-lihaoyi/Ammonite#893 should allow to handle more of the completion logic from almond, and make it interruptible in particular.