I wrote:
> The direct cause of the problem is that EvalPlanQual isn't completely
> initializing the estate that it sets up for re-evaluating the plan.
> In particular it's not filling in es_result_relations and
> es_num_result_relations, which need to be set up if the top plan node
> is an Append. (That's probably my fault.) But there are a bunch of
> other fields that it's failing to copy, too.
I believe I have fixed this problem in CVS sources for current and
REL7_1, at least to the extent that EvalPlanQual processing produces
the right answers for updates/deletes in inheritance trees.
However, EvalPlanQual still leaks more memory than suits me ---
auxiliary memory allocated by the plan nodes is not recovered.
I think the correct way to implement it would be to create a new
memory context for each level of EvalPlanQual execution and use
that context as the "per-query context" for the sub-query. The
whole context (including the copied plan) would be freed at the
end of the sub-query. The notion of a stack of currently-unused
epqstate nodes would go away.
This would mean a few more cycles per tuple to copy the plan tree over
again each time, but I think that's pretty trivial compared to the plan
startup/shutdown costs that we incur anyway. Besides, I have hopes of
making plan trees read-only whenever we do the fabled querytree
redesign, so the cost will someday go away.
Comments, objections?
regards, tom lane
