Fairness across multiple streams

[Lupus]

We ran into an issue with fairness when running multiple echo streams on one httpaf-based service. This problem boils down to one stream hogging all CPU power as reads never block on high data rates and event loop does not have a chance to run and schedule others. See ocsigen/lwt#622 for more context.

In our custom Lwt adapter we workaround this by adding Lwt_main.yield () after each read loop iteration along with increase of read buffer size to one megabyte, this way we yield once per each megabyte of data read, which gives good fairness and does not kill performance by endless epolls.

May be similar logic could be baked into Lwt adapter behind some configuration parameters?

[Lupus]

Actually after back-pressure feature has landed, adding Lwt.pause() inside on_read callback right before scheduling another read callback works as well as our forked Lwt adapter with vanilla httpaf, so we would probably settle on this approach, as it allows us to avoid patching&rebasing gluten upon each release.

[anmonteiro]

@Lupus I'm tempted to close this issue as a "won't fix", and instead recommend folks to use Lwt.pause inside their on_read handlers every N reads, like you do in your applications. What do you think?

[Lupus]

Yeah, sounds good to me!