Language server caching over-disposes LS instances, breaks on interpreter change

#8815 introduced LS caching and ref counting, such that the interactive window can reference the same language server as the regular code and keep it around until it’s no longer needed. The keys are resource scoped, which are good for ensuring things are shared and the minimum number of LSs are spawned.

Unfortunately, the activeServer and resource properties are essentially “globals” to the extension as a whole. The current code kills LSs as the user moves around different files to save resources, however this breaks multi-root workspaces which rely on being able to navigate around and still get code completions. (e.g., I open two files, one in each workspace, and switch between them)

Additionally, I believe the current code over-disposes, leading to a scenario where on the first active LS change, the next spawned server cannot be killed and isn’t configured correctly, leading both to https://github.com/microsoft/vscode-python/issues/5132#issuecomment-603571156 and a related bug where the next-spawned server’s interpreter doesn’t change.

The solution here is to eliminate activeServer and resource and make LSs fully resource-scoped (only accessible by key). Once a LS is no longer referenced (no handles from the interactive window, and its workspace no longer exists), the LS can be thrown away.

[nfrasser]

Has anyone tried the new Pylance extension? So far it seems to be working with multi-root workspaces for me.

[ErwanDL]

Are there any advancements on this bug ? This is actually a major painpoint for me, as it prevents me from being able to use MLPS on my main Python project at work (which uses multi-root workspaces).

Unless you deem this is a really advanced bugfix, I would love to help on it. Any guidance on where to start working ? 😃