On Fri, Sep 26, 2008 at 01:29:41PM +0200, Marcus Herou wrote:
> I've now launched HAProxy in production for the hardcore services :)
Great news !
> It's been up since yesterday about 16.00 +1GMT and pingdom reported some
> outages but I don't trust that too much since I have set the timeouts to a
> really low value in all servers to avoid connection starvation.
your site seems to respond quickly. In the first days, you might want to keep an eye on the logs, and check for connection timeouts and response timeouts (flags "sC" and "sH"). They always indicate some inappropriate configuration somewhere, or software problems which need to be worked around by increasing timeouts (especially for sH).
> One thing though I've noticed is that we have quite a lot of TIME_WAIT
> sockets (which I guess is expected in a LB), currently 12638. Is that
> totally normal and OK or should I tune the sysctl to drop them faster?
No, you don't have to do anything at that level, because TIME_WAIT sockets don't hurt at all. The real annoyance comes when you have to run netstat, but 12k sockets will pass quickly anyway. It's more annoying when you start getting 500k or 1M entries.
I also suggest that you enable the stats page (possibly on a different port) so that you can regularly monitor activity and get a synthetic report of errors and overall system health. That way you'll detect if the health check timeouts are too short for instance.
Willy Received on 2008/09/28 07:56
This archive was generated by hypermail 2.2.0 : 2008/09/28 08:01 CEST