Hi Marco, On Mon, May 18, 2015 at 08:27:25AM +0200, Marco Corte wrote: > I upgraded from version 1.5.11. > For these hosts I always use the ubuntu packages coming from Vincent > Bernat: https://launchpad.net/~vbernat/+archive/ubuntu/haproxy-1.5
I don't see any relevant change between 1.5.11 and 1.5.12 in this area. > I really did nothing else than a package update (apt-get dist-upgrade) > that also updated the kernel and a few other packages. The previous > update was about 3 weeks before. > I only saw a higher CPU load at first, then I correlate it to the higher > NIC traffic. > Please note that there is no real problem, since theservers still have > plenty of resources to use. OK but you're perfectly right to report abnormal observations before they become a concern. I'll try to reproduce here. In the mean time if you could double-check with 1.5.11 that the load really drops, that could be useful. > The only symptom that let me think to an haproxy issue was the traffic > and CPU usage drop in both peers, when turning off the _inactive_ > haproxy instance. Hmmm, that's interesting. What could be possible is that for any reason haproxy disconnects and reconnects all the time, transfering everything each time. > Noone is experiencing a similar behaviour, so I am starting thinking to > a problem in another package (the kernel, a library, ...). Or nobody noticed yet :-) > Any suggestion in how to investigate this further? I think the easiest thing to start with is to run "netstat -atn" on the backup node to verify if the peers connection is always between the same two ports or if it changes, indicating a reconnection. > I am already planning a reboot with an older kernel in the next days. I don't see how the kernel could impact the peers traffic alone and not the rest of the traffic. It is possible that one apparently harmless change or fix had an impact, but then we'll have to find which one. Note, there was a backported fix about a wrong timeout being applied, maybe it causes low-traffic sessions to expire once in a while and reconnect. I haven't observed this behaviour but that could be possible. Regards, Willy
