Reduce initial code intelligence latency

[beyang]

There's often an initial delay in getting code intelligence when a user refreshes a page. For extremely large repos, it can be many seconds, but even for smaller ones, a few seconds of noticeable wait time is often the case. For example, go to a random repository and see how long it takes the initial hover to become visible.

Is there anything we can do for 3.0 that either improves this across the board or for specific languages?

[sqs]

Basic code intel helps with this in many cases (and then hides when the precise language extension is ready).

For 3.0, I am fixing #1939 and #1938, which will fix issues in that interaction that occur for TypeScript. I want to get to a point where everyone always has basic code intel enabled and it basically never makes things worse. I believe we can get there by 3.0. Do you have any other ideas on how to improve this that we could do by 3.0?

[keegancsmith]

Is this due to not having a proxy ensuring you reconnect to an existing LS for a workspace?

[beyang]

@keegancsmith Yes, I think at least partly.

The idea I had (which would have to be implemented for each LS) was to reuse LS state over multiple connections, so that if 2 LSP connections have the same init parameters, they should use the same compiler/build state.

[chrismwendt]

The connection could be shared across browser tabs if there's only 1 instance of each extension running in the background script in the browser extension (do we already have an issue for this?).

go-langserver could reuse the same internal struct until a client issues a textDocument/didChange LSP notification, at which point it needs to clone the struct and update refcounts. That sounds a bit tricky, but could be simplified if go-langserver assumed all connections are stateless after initialize, which is true for the sourcegraph-go extension.

Backlogging this because:

• Basic code intel is on by default now https://github.com/sourcegraph/sourcegraph/issues/2035
• Time spent improving basic code intel will probably have a bigger impact on our Q1 goal than time spent on precise code intel

[chrismwendt]

One quick way to improve the perceived delay in code intelligence is to begin initialization of the language server immediately upon opening the file, rather than waiting for the first hover.

[chrismwendt]

The cache improvement PR has been merged and deployed sourcegraph/go-langserver#363

Time to first j2d should be faster when the zip archive has already been fetched earlier when a different user visited that repo@commit (go-langserver will reuse that same archive on disk instead of re-fetching the archive).