Re: Dealing with a server going down

On Sun, Nov 18, 2007 at 11:33:22PM -0500, Martin Goldman wrote:
> Hi folks,
>
> I'm new to haproxy, and I'm working on a simple SMTP cluster with 2 servers
> to learn how it works. I've got a quick question:
>
> When both SMTP servers are up and running, when I telnet to the cluster on
> port 25, I'm immediately connected and receive the SMTP server's greeting
> banner, sometimes indicating I'm connected to one server and sometimes to
> the other (demonstrating that load balancing is working). If I stop the SMTP
> service on one server and retry my connection to the cluster, I still get
> immediately connected and receive the SMTP greeting banner, now always
> indicating that I'm connected to the other server. But if I take one server
> off the network entirely and connect to the cluster once again, there's very
> often a delay before the SMTP greeting banner appears (the duration of which
> depends upon the value of contimeout). I suppose this means that haproxy is
> trying to send some of my requests to the server that's down, then it's
> waiting for the connections to time out before trying the other server.
>
> Now, I could be wrong, but it seems to me that this arrangement doesn't
> provide the highest quality of service when one of the nodes in your cluster
> goes down -- although all requests are completed eventually, some are sent
> to the bad node and have to time out before they can be serviced by another
> node. It seems like it should be possible to determine when a node goes
> down, and to have haproxy stop sending requests to that node until it comes
> back up.
>
> So I guess my question is: is there a way to do this? Or am I completely
> misunderstanding something?

I think that you have not enabled health checks on your servers. You should use the "check inter XXX" arguments on each server line. Also, keep in mind that there is an "smtpchk" option to perform SMTP checks instead of pure TCP checks.

Regards,
Willy Received on 2007/11/19 11:58