There is a max sleep-timeout in each eventer. We have articifially high latency when we do asynch dispatch for single ops.
- HTTP request arrives (eventer fires)
- request read
- event descheduled and asynch job dispatched.
- job finished and backq's result.
- no fd events fire, so poll timeout is reached (200ms)
- backq is consumed, eventer triggered from continuation job and HTTP response sent.
Proposed fix... if events fire, poll timeout is dropped to 0. For each cycle though the eventer that poll times out, add 1ms (up to eventer implementation max), any activity resets to 0.