Excerpts from Peter Eisentraut's message of dom ene 15 10:00:03 -0300 2012:
>
> On ons, 2011-12-28 at 14:35 -0500, Tom Lane wrote:
> > Alvaro Herrera <alvherre@commandprompt.com> writes:
> > > Excerpts from Peter Eisentraut's message of mié dic 28 15:04:09 -0300 2011:
> > >> On a hot standby, this fails with:
> > >> ERROR: cannot assign TransactionIds during recovery
> >
> > > I think we could just have the xid_age call
> > > GetCurrentTransactionIdIfAny, and if that returns InvalidXid, use
> > > ReadNewTransactionId instead. That xid_age assigns a transaction seems
> > > more of an accident than really intended.
> >
> > The trouble with using ReadNewTransactionId is that it makes the results
> > volatile, not stable as the function is declared to be.
>
> Could we alleviate that problem with some caching within the function?
Maybe if we have it be invalidated at transaction end, that could work.
So each new transaction would get a fresh value. If you had a long
running transaction the cached value would get behind, but maybe this is
not a problem or we could design some protection against it.
For the check_postgres case I imagine it opens a new connection on each
round so this would not be a problem.
--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support