[ejabberd] Amazon load balancer and ejabberd, problematic?

Keith Irwin keith.irwin at gmail.com
Sun Apr 10 01:08:18 MSD 2011


On Apr 6, 2011, at 6:45 AM, Armando Di Cianno wrote:

>> My question isn't about the JVM stuff (quite yet) but about whether or not
>> there's something odd about the behavior of running XMPP through a load
>> balancer (Amazon's ELB, specifically) that can occasionally break things in
>> a way that's not really normal for XMPP streams.
>> Any wisdom out there?
> 
> I've had various levels of luck with ejabberd and load balancers; I
> have not used ELB.
> 
> Most load balancers worked just fine, with differences in performance.
> I had some issues with HAProxy 1.13.x (haven't tried 1.14.x), where no
> amount of configuration of timeouts (client->haproxy, haproxy->server,
> or full client->server) seemed to allow the connection to remain up
> reliably. You may be having a similar situation -- while the balancer
> supports generic TCP connections, it may be inherently for or
> tuned-for mostly short lived HTTP connections.
> 

I finally found a reference somewhere to the fact that ELB (Elastic Load Balancer) has a 60 second timeout if there's no data flowing through. The library I'm using sends keep-alives every 30 seconds, so I'm relatively safe. Alas, that library had another bug which made it impossible to reconnect if you get disconnected while in a chat room.

Anyway, thanks for the info!

Keith



More information about the ejabberd mailing list