On Fri, Aug 10, 2012 at 7:19 AM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
Haifeng Liu <liuhaifeng@live.com> writes: > I have a program running like a daemon, which analyze data and write to postgresql 9.1 on centos 5.8. There is only one connection between my program and the postgresql database, and I hope the connection may keep alive all the time. But I failed, the connection will be reset after idle for about 2 hours.
> jdbc driver: 9.1-901, connection url has parameter tcpKeepAlive=true; > postgresql:9.1, keep alive related settings use default values(commented); > centos 5.8 64bit, net.ipv4.tcp_keepalive_intvl = 75, probes = 9, time = 7200.
IIRC, time = 7200 (seconds) means to start sending keepalive packets after 2 hours of idle time. So if you have something in the way that is dropping the connection after 2 hours, these settings will not activate keepalive soon enough to save it. I'd try setting that to 3600.
I use a much shorter time (300). Why not? The keepalive pings are a trivial amount of data. I've found over the years that router NAT tables are a big offender (probably doesn't apply in this case) ... some drop your internal-to-external IP address mapping after as little as five minutes.
Craig
> There is no firewall or any other device which behaves force idle connection cleanup.
Seems pretty darn unlikely given these symptoms. Look harder... maybe something you thought was just a bridge has got routing behavior.
BTW, in the real world connections drop for all sorts of reasons, and kernel keepalive configurations can't prevent them all. You might be better advised to build some reconnect-after-connection-loss logic into your application, rather than spending time on making this work.
