On Tue, 2010-10-26 at 16:08 -0400, Robert Haas wrote:
> On Tue, Oct 26, 2010 at 8:10 AM, Simon Riggs <simon@2ndquadrant.com> wrote:
> > I agree with your analysis. Let me review...
> > [review]
>
> Sounds like we're on the same page.
>
> > Two options for resolution are
> >
> > 1) Throw SERIALIZABLE error
> >
> > 2) Implement something similar to EvalPlanQual
> > As you say, we already resolve this situation for concurrent updates by
> > following the update chain from a row that is visible to the latest row.
> > For MERGE the semantics need to be subtely different here: we need to
> > follow the update chain to the latest row, but from a row that we
> > *can't* see.
> >
> > MERGE is still very useful without the need for 2), though fails in some
> > cases for concurrent use. The failure rate would increase as the number
> > of concurrent MERGErs and/or number of rows in source table. Those
> > errors are no more serious than are possible now.
> >
> > So IMHO we should implement 1) now and come back later to implement 2).
> > Given right now there may be other issues with 2) it seems unsafe to
> > rely on the assumption that we'll fix them by end of release.
>
> Yeah. In fact, I'm not sure we're ever going to want to implement #2
> - I think that needs more study to determine whether there's even
> something there that makes sense to implement at all. But certainly I
> wouldn't count on it happening for 9.1.
2) sounds weird, until you realise it is exactly how you would need to
code a PL/pgSQL procedure to do the equivalent of MERGE. Not doing it
just makes no sense in the longer term. I agree it will take a while to
think it through in sufficient detail.
In the meantime it's a good argument for the ELSE construct at the end
of the WHEN clauses, so we can do something other than skip a row or
throw an ERROR.
-- Simon Riggs http://www.2ndQuadrant.com/books/PostgreSQL Development, 24x7 Support, Training and Services