I've been chasing Chris Bitmead's coredump report from earlier today.
I find that it can be reproduced very easily. For example:
regression=> select f1 from int4_tbl group by f1;
< no problem >
regression=> select f1 from int4_tbl* group by f1;
< core dump >
(You may get unstable behavior rather than a reliable core dump
if you are not configured --enable-cassert.)
The problem seems to be in optimizer/plan/planner.c, which is
responsible for creating the Sort and Group plan nodes needed to
implement GROUP BY. It also has to mark the lower plan's targetlist
items with resdom->reskey numbers so that the executor will know which
items to use for sort keys (cf. FormSortKeys in executor/nodeSort.c).
The trouble is that that latter marking is done in planner.c's
make_subplanTargetList(), which *is never invoked* for a query that
involves inheritance. union_planner() only calls it if the given plan
involves neither UNION nor inheritance. In the UNION case, recursion
into union_planner does the right thing, but not so in the inheritance
case.
I rewrote some of this code a couple months ago, but I find that 6.4.2
has similar problems, so at least I can say I didn't break it ;-).
It seems clear that at least some of the processing that union_planner
does in the simple case (the "else" part of its first big if-then-else)
also needs to be done in the inheritance case (and perhaps also in
the UNION case?). But I'm not sure exactly what. There's a lot going
on in this chunk of code, and I don't understand very much of it.
I could really use some advice...
regards, tom lane
