Fork and adapt GitLab Workhorse for Mercurial
Gitlab Workhorse is a reverse proxy sitting between Nginx and the Rails application.
Among its purposes is to deflect any Git HTTP requests that are potentially long (pull, push) directly to Git without going through the Rails application. Among its obvious performance and robustness benefits, this helps keeping the unicorn (Rails server) to a minimum, while Workhorse itself is designed to handle long timeouts gracefully. I suppose it's also able to stream requests and responses completely, something I doubt the Rails app would do easily.
As far as I know, it works by first calling an internal API on the Rails application for authentication and authorization, calling the inner Git serving process depending on the result.
We would need to have the same for Mercurial. Actually going through the Rails app has already been a problem in the past, because of weird handling of
Content-Encoding: chunked when we switched from
hg serve to
gunicorn. Obviously, it would be much faster and easier on the RAM than what we're currently doing (reading the whole from Mercurial, I think that actually de-chunks IIRC).
That means probably to fork Workhorse. Perhaps we'd have to wait until we get back to the standard Omnibus build/deploy system. Or maybe we'll want that so much at that point that we'll be happy to build from
heptapod-docker. We'll see.