Kernel Providers should be deprecated in jupyter client

[kevin-bates]

While looking into how nb_conda_kernels work I realized that jupyter_client has discovery.py - which exposes the KernelProvider work that is really taking place in @takluyver's repos: takluyver/jupyter_kernel_mgmt and takluyver/jupyter_protocol. In addition, I see that there are folks using the kernel provider framework in jupyter_client - which includes the make_manager() method - which no longer exists in the definitive framework.

The plan is that the jupyter_server will leverage jupyter_kernel_mgmt's KernelProvider framework. The framework will support asyncio and parameterized launches (aka parameterized kernels), etc. There are no plans to use this framework from within jupyter_client. Applications using jupyter_client will need to switch to using juptyer_kernel_mgmt/juptyer_protocol to take advantage of KernelProviders, async management and parameterized launches. This does not preclude building similar support into jupyter_client's kernel management - although those efforts should not include the KernelProvider architecture.

As a result, we should deprecate jupyter_client's discovery.py, and the classes that it implements, with proper messaging to use jupyter_kernel_mgmt instead. Fortunately, discovery.py is not currently included in any releases, nor should it be.

[SylvainCorlay]

Hi @kevin-bates I think that we can close the issue as we are going to point master at the current head of 5.x.

[kevin-bates]

I agree - thanks @SylvainCorlay