Thread: plpython trigger code is wrong (Re: [GENERAL] Potential bug -- script that drops postgres server)

Lance Thomas <LThomas@DevIS.com> writes:
> Below is something that may be of interest -- a short, 7-statement script
> that seems to drop my postgres server.

It appears that the plpython trigger implementation assumes that any
given procedure will be used as a trigger for only one relation.  The
reason it crashes is it's trying to use the rowtype info of the relation
it was first compiled for with the other relation.

Probably the easiest fix is to include the relation OID as part of the
Python name of a trigger procedure, so that a separate copy is compiled
for each relation the procedure is used with.

Any plpython users want to step forward and fix this?  I have other
things on my plate ...

            regards, tom lane

PS: I haven't tested, but I wonder whether any of our other PLs have the
same bug.

* Tom Lane (tgl@sss.pgh.pa.us) [030214 19:35]:
> Lance Thomas <LThomas@DevIS.com> writes:
> > Below is something that may be of interest -- a short, 7-statement script
> > that seems to drop my postgres server.
>
> It appears that the plpython trigger implementation assumes that any
> given procedure will be used as a trigger for only one relation.  The
> reason it crashes is it's trying to use the rowtype info of the relation
> it was first compiled for with the other relation.
>
> Probably the easiest fix is to include the relation OID as part of the
> Python name of a trigger procedure, so that a separate copy is compiled
> for each relation the procedure is used with.

Interesting idea.  I had been taking the approach of applying the cache
to just the python compilation, and not the rowtype info.  This has a
substantial performance penalty, which I'd been addressing by eliminating
some unneeded parameter processing that doesn't apply in the trigger
case, and considering a separate cache for each rowtype.

> Any plpython users want to step forward and fix this?  I have other
> things on my plate ...

I'm looking at the bug right now.  Patch in a day or so.

-Brad