Tom Lane wrote:
> With the fix I was just about to apply, all four cases give the first
> set of results. This clearly satisfies the principle of least
> astonishment, at least more nearly than what we have; but it equally
> clearly is *not* going to restore 8.4 to work just like 8.3.
Right, 8.3 had the same underlying problem, 8.4 just makes it more
visible as it's better at flattening subqueries.
> I'm inclined to apply the patch to 8.4 anyway, because it seems like a
> bug fix. I would consider patching further back except there's no
> chance of making it work in older branches, at least not without
> destabilizing them quite a bit (the PlaceHolderVar mechanism would have
> to be back-ported).
>
> It might be possible to fix the older branches by not flattening
> subqueries that have whole-row references; but even that would take
> nontrivial work, and it'd be sacrificing performance to fix a corner
> case no one has previously complained about. So I'm leaning to patching
> 8.4 and leaving the older branches alone.
>
> Thoughts?
Seems reasonable.
--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com
