On Tue, 2007-05-08 at 15:59, Sean Murphy wrote:
> Scott Marlowe wrote:
> > On Tue, 2007-05-08 at 12:19, Sean Murphy wrote:
> >> Tom Lane wrote:
> >>> Sean Murphy <Sean.Murphy@equipoint.com> writes:
> >>>> I'm WAY out of my depth here, but my impression, based on the
> >>>> circumstances, is that there is some sort of an idle session timeout
> >>>> kicking in (most likely on the client side) and dropping the connection.
> >>> There's no such timeout in the Postgres server, for sure. I would
> >>> actually bet that your problem is in some router between the client and
> >>> the server. In particular, routers that do NAT address mapping
> >>> typically have a timeout after which they will forget the mapping for an
> >>> idle connection. If you've got one of those, see if it'll let you
> >>> change the timeout.
> >>>
> >>> If you can't do that, you might think about teaching your client-side
> >>> code to send dummy queries every so often.
> >>>
> >>> regards, tom lane
> >>>
> >> I've already maxed out the connection timeout at the firewall... and
> >> I've been using dummy queries every five minutes, but it just feels like
> >> a crutch to do so.
> >
> > Have you looked into tcp keepalive settings?
> >
> > net.ipv4.tcp_keepalive_intvl = 75
> > net.ipv4.tcp_keepalive_probes = 9
> > net.ipv4.tcp_keepalive_time = 500
> >
> > Not sure if those settings will help with an NAT router or not but it's
> > worth a try. Pgsql 8.2 can set those for you.
> >
>
> I may be celebrating prematurely,
Never stopped me :)
> but resetting the tcp_keepalive
> parameters seems to have done the trick - I left a pgAdmin connection
> that *always* drops after inactivity up while I went to lunch and it was
> still alive when I got back.
>
> Is there a way to alter the tcp_keepalive settings on an app-by-app
> basis rather than for the whole system?
Well, you could set it on individual workstations instead of on the
server. I.e. if you set tcp_keepalive on your workstation to 500, but
leave Wally and Dilbert set at the default 7200 then they'd still
timeout and you wouldn't.