Thoughts on backpressure

[tower-rs]

Thoughts on backpressure

[carllerche]

tower-balance is the counter example.

In this case, tower-balance holds a set of inner services. When Balance::poll_ready is called, it in turn calls poll_ready on all the inner services. When a request is received, it is dispatched to a ready inner service.

The problem is that poll_ready has been called on all the inner services, which reserves capacity for a request. A request is only sent to one inner service. All the other inner services then have capacity for one request held in limbo, unable to be used for any other service clones. The capacity is not released until those other services are selected to receive requests.

Thoughts @olix0r

[jonhoo]

The proposed and_then solution smells a lot like it could introduce deadlock? Imagine two threads, 1 and 2, which try to call two service in opposite orders. They both reserve the last (/only) slot in their "first" service, and will now wait forever to acquire a slot in the second.

[carllerche]

@jonhoo The existing (and really any concurrent code) is at risk of introducing deadlocks. Granted, this makes it a bit easier.

Given that the only other option (that I can think of) is unbounded buffering, I think it is a necessary hazard.

[jonhoo]

Yeah, I mostly raised it because this seems like a pre-made deadlock trap, much like locks are, and should probably be documented as such. That's not to say I think we shouldn't have it — we should just make sure that users are aware that this can deadlock.

[carllerche]

This is implemented, but we need to document it better.

[carllerche]

Closing in favor of the doc meta issue (#33).