Enable incrementally updating diagnostics about language-server-protocol HOT 3 OPEN

The pull diagnostics method as an unchanged property to signal that issue for a document have not changed. Incremental updating does not exist right now.

from language-server-protocol.

@keyboardDrummer Can you give us a more detailed overview what you think is a "large file" and awhat a lot of diagnostics would be?

Because in my experience the reason why things get slower is not the fact that a 100-1000 diagnostics are sent from the server to the client, but the fact that the server is busy parsing and doing type inference (which is probably 100 times more time intensive than sending the data).

from language-server-protocol.

Consumers of our VSCode extension (for Dafny) have files that are thousands of lines, and information diagnostics are likely to occur on every line. An information diagnostic per line may seem like a lot, but our language has a relatively high level of implicitness/inference (unrelated to types), so it can help to tell the user what the compiler has inferred from their code.

Sending a 1000 diagnostics for a single file might be OK, but we have the additional situation that we may send many publishDocument notifications for a single document change.

The reason for this is that our back-end has one fast phase (parsing/name resolution/type inference) that takes about a 100ms, but then does a complicated analysis (called verification) on separate program elements, of which a large file may have hundreds, that each can take between 100 and 1000 ms. Automatically running verification is in some ways very similar to automatically running unit tests. We publish diagnostics after the fast phase, but also each time verification of any element completes, so we may send a 100 publishDocument notifications.

What I'm worried about is sending 1000 (#lines * diagnostics/line) * 100 (amount of publishDocument calls) of elements per document change. I'll admit we are not evidently running into this at the moment, but it seems like our current design won't scale towards ever growing customers codebases.

I will say though we've extended LSP with a notification API to notify the client about the current progress of verification, so for each verifiable element to say whether it's: not being run/queued/running/completed. We could send additional diagnostics through this API, and use publishDocument only for the fast phase parsing/name resolution/type inference) diagnostics.

from language-server-protocol.