On Wed, Aug 25, 2010 at 09:26:51AM +0300, Heikki Linnakangas wrote:
> On 24/08/10 23:56, Andres Freund wrote:
> >I have to ask one question: On a short review of the discussion and
> >the patch I didn't find anything about the concurrency issues
> >involved (at least nodeModifyTable.c didnt show any).
> The SQL spec doesn't require MERGE to be an atomic "upsert" operation.
> >Whats the plan to go forward at that subject? I think the patch needs
> >to lock tables exclusively (the pg level, not access exclusive) as
> >long as there is no additional handling...
> Well, you can always do LOCK TABLE before calling MERGE if that's
> what you want, but I don't think doing that automatically would make
> people happy.
But randomly loosing tuples will make much more people unhappy. At a
much more problematic point of time (in production).
There is no locking prohibiting situations like trying to update a
tuple which was concurrently deleted and thus loosing a tuple. Unless
I miss something.
Andres