Tom Lane <tgl@sss.pgh.pa.us> writes:
> In an RFC-compliant stack, the outage interval required before KEEPALIVE
> will kill the connection is of the order of hours. RFC 1122 specifies
> that the minimum interval before the first probe is even sent is 2 hours
> (since last activity on connection), and that a single failed probe is
> not sufficient reason to drop the connection.
Ah, yes, I see that now. On NetBSD, it's four hours until the first
keepalive, then eight missed ones at 150 second intervals (totalling
20 minutes) are required before the connection is considered dead.
> RFC 2525 does note that excessively short keepalive timeout is a common
> form of TCP-stack bug.
So, Bruce might still be bothered with something like that, and/or
(for all he's given us of details) he might actually be talking about
a situation where Oracle will wait through severely prolonged outages
where PostgreSQL won't.
Details, Bruce?
-tih
--
Puritanism -- the haunting fear that someone, somewhere, may be happy.