Support refreshing single files in diagnostic pull

[rchiodo]

In this code here, workspace diagnostics is repolling every 2 seconds:

vscode-languageserver-node/client/src/common/diagnostic.ts

Line 440 in d90a87f

this.pullWorkspace();

This makes Pylance continuously do analysis. At least for items that don't seem to save the fact that there are no errors. (Previous result ids don't exist for files that aren't open, but workspace mode asks for all files).

I'm not sure if the polling is as design (might be) but missing the previous result id seems like a bug?

For example, in the issue where we found this:
microsoft/pylance-release#7499

The user has a single file open.

Our result contains all files though. The next poll request only has a previous result id for the file that's open. If all previous results were passed, we would mark the next request as unchanged, but we can't because of the missing result ids for the unopened files.

[dbaeumer]

@rchiodo the idea of workspace pull is that you as long as possible keep the request open (don't return a response and stream results on file changes into the request). Are you closing the request after every pull?

[rchiodo]

Oh, we just return as soon as we have all of the results. So, we should be using partial results?

That seems like push mode instead of pull.

[dbaeumer]

That seems like push mode instead of pull.

It is a long running pull. There is no indication on the client side to when trigger a new pull. That is why it is implemented that way.

[rchiodo]

The refresh request indicates to the client that a new pull is required? I guess we could just use the spots where we ask for a refresh to instead send more progress, but for workspace diagnostics, we'll essentially have the same logic we have for push mode.

[dbaeumer]

The difference to the old model is that the file and workspace pull are coordinated.

I am open for a feature that only pulls once and then waits for a refresh.

[rchiodo]

I've reimplemented our workspace diagnostics to use the partial results model and it seems better. We don't reanalyze everything on the refresh. Instead, we just send changes as they happen.

So, I'm not sure how it could be better. Maybe if there was a way to say 'refresh this file'? We know which ones need to have analysis redone, but maybe that's not the same for all language servers.

[parsa2000200]

In this code here, workspace diagnostics is repolling every 2 seconds:

vscode-languageserver-node/client/src/common/diagnostic.ts

Line 440 in d90a87f

this.pullWorkspace();

This makes Pylance continuously do analysis. At least for items that don't seem to save the fact that there are no errors. (Previous result ids don't exist for files that aren't open, but workspace mode asks for all files).

I'm not sure if the polling is as design (might be) but missing the previous result id seems like a bug?

For example, in the issue where we found this:
microsoft/pylance-release#7499

The user has a single file open.

Our result contains all files though. The next poll request only has a previous result id for the file that's open. If all previous results were passed, we would mark the next request as unchanged, but we can't because of the missing result ids for the unopened files.