On Mon, Aug 16, 2010 at 9:22 PM, Willy Tarreau <[email protected]> wrote: > Indeed, that looks strange. I don't remember about anything like this, > and I even remember being careful about not logging this state. Maybe > this has changed since this version, I don't know. Just one hint, do > you see any failed check on this server's stats (I just mean failed > checks, not down transitions) ?
I just realized we have Zabbix storing the HAProxy status CSV every minute; and what I'm seeing is that there's one whole minute when *all* backends on one particular box are suddenly marked as DOWN. At that exact same time, I see that the load average suddenly spiked at 35 -- so something weird is happening on that box (I hate weird behaviour). Anyway, the reason we are missing the DOWN messages must be that the syslog daemon was not getting the log packets. This is why I don't like syslog over UDP. :-(
