Kernel Finder workflow

[rebornix]

This is part of the architecture review and refactoring task #10832 and majorly focused on kernel finding/detection.

The current workflow of finding kernels (10,000 feet view) is

1. Extension activation / notebook document open
2. Find Kernels
   • Find kernels from cache
     • Find kernels from remote server globalState
     • Find kernels from local kernel globalState
     • Validate Kernels from cache (async)
   • Update NotebookControllers
   • Find kernels without cache
     • Find kernels from remote server, reach out to the server
     • Find Kernels from local
       • Searches for kernels on the file system in well known paths documented by Jupyter
         • If we already search for it once, use that result (unless it's empty)
       • Returns all Python kernels and any related kernels registered in the python environment
         • Get interpreters from Python
         • Find python global python kernel specs and non python specs
   • Update NotebookControllers
3. Listen to following events
   • Remote server Uri change
   • Remote kernel start/dispose
   • Extension change
   • Conda Environment change
   • Interpreters change
4. When receive above events, run step 2.

As we register all the event listeners at a very top level, we accidentally trigger redundant kernel fetching requests. For example:

• When a new kernel session is created on a remote server, we will run step 2, meaning we would try to find local python kernels without cache again.
• When interpreters change, we would run step 2 again and send requests to remote server to fetch kernels/sessions

Also, when we re-run step 2, we always read cache first (which is asynchronous and not-zero-cost as it needs to validate all kernel specs) and then invalidate all caches. This makes sense on activation, but it's not necessary on event handling.

[rebornix]

What we can do to improve the redundancy of kernel finding

• On extension activation, find from cache, and then find without cache
• On events
  • Update relevant caches
  • Broadcast kernels change event
  • ControllerLoader request to find from cache again

[DonJayamanne]

Done