requests sit in job queue for thread_job_max_queuing_milli_seconds before finally successful

### HHVM Version

HipHop VM 3.25.3 (rel)
Compiler: tags/HHVM-3.25.3-0-g4f0d713d5613da7f3a819392d4e1e7addb4f64f7
Repo schema: 850169ff35b7a904e388714613c366ac03c486ab

hh-df226aa5fe99a582256b4227c2e8ab2fa64a8c87-3.25.3 Fri May 4 15:29:09 2018

### Operating System and Version

Centos 7

##########

Under high traffic, we see bursts where several requests have a long duration.

For example, time measured within hack code is 3 ms, but time measured from nginx in front of hhvm is 2.003 s.

We have `hhvm.server.thread_job_max_queuing_milli_seconds = 2000`, which sort of explains the story, but the requests actually succeed after 2 seconds of being in the queue.

I'd think we perhaps have a LIFO/FIFO setting wrong somewhere, but we also dont specify `hhvm.server.thread_job_lifo_switch_threshold`, which makes me think that we're always in FIFO mode.

During the 2.003 seconds for that request (idle? in the job queue?), other requests are being handled successfully.

Is this a bug? Or maybe I just dont understand what's going on.

Thanks for your help!