Only cache ports if the cache_ports flag is set to True

[martinRenou]

cc. @kevin-bates @SylvainCorlay

[kevin-bates]

@martinRenou - thank you for adding the flag. Since MultiKernelManager can manage different types (and localities) of kernels, the decision to cache ports for a given KernelManager should, ideally, come from the KernelManager instance.

[SylvainCorlay]

@martinRenou - thank you for adding the flag. Since MultiKernelManager can manage different types (and localities) of kernels, the decision to cache ports for a given KernelManager should, ideally, come from the KernelManager instance.

This sounds like a convoluted design while we are not quite sure why disabling that behavior is required in the first place - and you can still disable it completely.

Would you like to have a quick chat to discuss this viva voce?

[kevin-bates]

@SylvainCorlay Sure - I think that would be great. We can meet here: https://ibm.webex.com/meet/kevin.bates unless you have something else.

[SylvainCorlay]

I am on the room.

[martinRenou]

As discussed during the meeting. We'll add the flag in KernelManager, in case the MultiKernelManager manages remote and local kernels (this caching logic only makes sense in the case of a local kernels).

[martinRenou]

Would you like to give it a look @kevin-bates ?

[kevin-bates]

Looks good @martinRenou - thank you!

[SylvainCorlay]

Go!

[meeseeksmachine]

This pull request has been mentioned on Jupyter Community Forum. There might be relevant details there:

https://discourse.jupyter.org/t/jupyter-client-6-0-0/3414/1

[kevin-bates]

@martinRenou - I took JC 6.0 for a spin in my EG environment yesterday and had an issue with cache_ports during shutdown because the cached ports weren't in currently_used_ports resulting in a KeyError.

Clearly, the remote kernels don't need (or want) to cache the ports, and I need to change cache_ports to false for these kinds of manager instances. However, I figured I'd be able to do this via the command line (e.g., --KernelManager.cache_ports = False) but now realize that the cache_ports traitlet isn't configurable. Moreover, for subclasses to set cache_ports = False, they must implement a constructor (or whatever __init__() is called in python) because the cached ports are captured immediately following the manager's instantiation. Once I created the constructor, I was able to manage the remote kernels w/o a KeyError.

My question is, was it intentional to not have the cache_ports traitlet configurable (I.e., config=True)? I'm curious how someone would set it to false if they still wanted tcp without having to modify their code, create a new release, etc.? I think being able to do something like --XXXKernelManager=False while still allowing KernelManager instances to cache ports would be nice.

I may not fully understand how traitlets work, so wanted to get your take on how users should change cache_ports.

cc: @SylvainCorlay