fair load balancing?

[Moses Hohman]

On Jun 23, 2008, at 2:50 PM, dormando wrote:

>>
> Backend goes back to the pool. Options request is sent once per
> established backend connection; so if you connect to a backend and  
> reuse
> it 500 times, you'd have one OPTIONS request and 500 normal requests.


And, because of this, I don't think the proposed solution works for  
the "long-running mongrel request" problem with persist_backend on.  
Also, if the OPTIONS request actually is served by the rails stack,  
then the OPTIONS request hangs just like a normal request would, and  
you see the same undesirable behavior (a normally short request taking  
a long time because it's waiting for a response from a mongrel that is  
serving a long-running Rails request). I think in order for this to  
work as intended you need both persist_backend off (which does mean  
double the number of requests to your mongrels, but probably that's  
not a huge deal) and you need to use the mongrel_rails option --num- 
processors 1, which will limit the number of worker threads operating  
in the mongrel to 1 at a time. Even when I do this, though, I'm still  
able to get the web stack to hang without too much effort by just  
reloading a page many times in succession very quickly (albeit with a  
very fast network connection to the server). It's entirely possible  
that I'm still missing something here, though. By chance I'm going to  
get to talk to Adam in person tonight, so I hope to sort these details  
out with him then, and I will report back to the list.

I think a while back in this thread someone said this was not a  
Perlbal problem, but a mongrel problem, but it would be more accurate  
to say this is a Rails problem. Mongrel can handle concurrent requests  
just fine. Rails imposes the single-thread of execution mutex lock.

This could also be viewed as an application-level problem (i.e. why  
are your requests taking longer than a couple seconds anyway, and if  
they must, why not offload the actual processing to a background job  
and have the web browser poll until the background job is done), which  
is sort of true. However, it would be easier if the load balancer  
could just handle this for us, as that would also insulate us from  
unanticipated slow requests as well.

Moses