On Fri, 2005-04-22 at 14:43 -0400, Alvaro Herrera wrote:
> On Fri, Apr 22, 2005 at 02:24:57PM -0400, Jan Wieck wrote:
> > On 4/22/2005 2:08 PM, Tom Lane wrote:
> >
> > >Sven Willenberger <sven@dmv.com> writes:
> > >>We have a replication set up between 2 servers using Slony; both are
> > >>runnind PostgreSQL 8.0.1. The issue is that when updates/inserts are
> > >>made to a replicated table, the replication does not occur; apparently
> > >>this is due to spi_exec somehow not allowing/causing the slony trigger
> > >>function to fire.
> > >
> > >Yuck :-(. The only idea that comes to mind is that 8.0 changed the
> > >timing of trigger firing --- the triggers are probably firing while your
> > >function still has control, whereas in earlier releases they'd only fire
> > >after it returns. Could this be breaking some assumption Slony makes
> > >about the order of operations?
> >
> > Slony triggers are AFTER ROW triggers. All they do is one SPI_execp() to
> > insert the log row. The only way that could possibly be suppressed is by
> > bypassing the executor and doing direct heap_ access.
> >
> > So how does plperl manage that?
>
> It doesn't; it only uses SPI. I guess we would need the original
> function to see what is really happening.
>
If by the "original function" you mean the plperl function, it was in
the thread parent; in essense it runs an "insert into tablename (cols)
select vals from <other table>" query. The Slony trigger is:
"_T2_logtrigger_10" AFTER INSERT OR DELETE OR UPDATE ON inter_rootdecks
FOR EACH ROW EXECUTE PROCEDURE "_T2".logtrigger('_T2', '10', 'kkkv')
If by "orginal function" you mean the logtrigger (slony) function it
would appear from a cursory glance at the source that logtrigger
accesses the heap (HeapTuple). It would appear the same manipulation is
utilized by plperl's SPI calls. How this all interacts with the
"executor" is, admittedly, beyond the scope of my comprehension at this
point ...
Sven