Andres Freund <andres@2ndquadrant.com> writes:
> On Thursday, October 04, 2012 12:16:14 AM Daniel Farina wrote:
>> I suppose this might needlessly eliminate someone who forks and hands
>> off the PGconn struct to exactly one child, but it's hard to argue
>> with its simplicity and portability of mechanism.
> Even that scenario isn't easy to get right... You need to get the socket from
> libpq in the parent after the fork() and close it. Only after that you can
> PQfinish it. Which you probably need to do before establishing other
> connections.
No, it's much easier than that: the parent can simply forget that it has
a PGconn. It will leak the memory occupied by the PGconn object, and it
will leak an open socket (which will only be half-open after the child
does PQfinish). This would be noticeable if the parent is long-lived
and creates many such connections over its lifespan, but otherwise
people could be doing it just fine. In fact, I had to look closely to
convince myself that pgbench didn't do it already.
I suspect that if we provide a mechanism like this, we'll have to
provide a way to turn it off, or somebody is going to complain that
we broke their code.
regards, tom lane