Lags issue with sinatra thin app - ruby

Question

I have some strange lags issue with a sinatra thin app when the trafic get a little high (30-40 users).
It's a small game using long-polling so http IOs can be high compared to the number of users.
CPU load stay low and there's a lot of free memory.
Here are some typical logs lines when lags happen :
1 - [17/Jul/2015:16:50:17 -0400] "POST /play?next=word HTTP/1.1" 200 1 0.0018
2 - [17/Jul/2015:16:50:17 -0400] "GET /update?_=1437166100579 HTTP/1.1" 200 304 15.0046
3 - [17/Jul/2015:16:50:17 -0400] "GET /update?_=1437166102348 HTTP/1.1" 200 286 15.0045
4 - [17/Jul/2015:16:50:17 -0400] "POST /accept_replay? HTTP/1.1" 200 - 0.0021
5 - [17/Jul/2015:16:50:18 -0400] "GET /core HTTP/1.1" 200 3719 0.0015
6 - [17/Jul/2015:16:50:18 -0400] "GET /join HTTP/1.1" 302 - 0.0640
7 - [17/Jul/2015:16:50:18 -0400] "GET /core HTTP/1.1" 200 3719 0.0024
8 - [17/Jul/2015:16:50:19 -0400] "POST /play?next=word HTTP/1.1" 200 1 0.0034
9 - [17/Jul/2015:16:50:19 -0400] "GET /update?_=1437166215907 HTTP/1.1" 200 248 10.0018
10- [17/Jul/2015:16:50:19 -0400] "GET /update?_=1437166222579 HTTP/1.1" 200 252 11.0029
11- [17/Jul/2015:16:50:31 -0400] "GET /core HTTP/1.1" 200 3719 0.0034
12- [17/Jul/2015:16:50:31 -0400] "POST /sentiment/bad? HTTP/1.1" 200 - 0.0024
13- [17/Jul/2015:16:50:31 -0400] "GET / HTTP/1.1" 200 4449 0.0086
14- [17/Jul/2015:16:50:31 -0400] "POST /decline_replay HTTP/1.1" 302 - 0.0020
And 30 more exactly at [17/Jul/2015:16:50:31 -0400]
( get /update are longpolling requests so it can take up to 40seconds)
Everythings stops for 12seconds between 10 and 11. And all the requests received during this time seems to be processed simultaneously.
I start the app that way
thin start -p 80
Can it be a thin issue ?
Do i need a custom thin config file ?
Do i need nginx?
Any indication is welcome...
edit :
Errors I find in ObectSpace [SystemStackError, 1][NoMemoryError, 1][IOError, 1]

Answer 1

The kind of behavior smells a lot like request queuing, which means there aren't enough web processes free to handle incoming requests. So the requests sit waiting, and then then the backlog is cleared they suddenly all get processed super fast and all at once.
This guy wrote up a good post on how to use Thin, EventMachine, and Async Sinatra to handle long-polling requests.