ERROR: MergeAppend child's targetlist doesn't match MergeAppend
Hi!
I ran into a problem with PG 9.1 and bug is observed even in master. After
simplifying a query (original was 9Kb long!) it's possible to reproduce it easily:
CREATE TABLE wow (t1 text, t2 text);
CREATE INDEX idx ON wow (t1,t2);
SET enable_seqscan=off;
SET enable_bitmapscan=off;
EXPLAIN
SELECT
t1, t2
FROM (
SELECT t1, t2 FROM wow
UNION ALL
SELECT 'a', 'a' FROM wow
) t
ORDER BY t1, t2;
if second 'a' constant is changed to something else then it works fine.
The root of problem is that tlist_member() (called in
create_merge_append_plan()) for second constant returns TargetEntry for first
constant because they are equal. And the same problem is observed if second
select is replaced by "SELECT t1, t1 FROM wow".
It's seems to me that check in create_merge_append_plan() is too restrictive:
if (memcmp(sortColIdx, node->sortColIdx,
numsortkeys * sizeof(AttrNumber)) != 0)
elog(ERROR, "MergeAppend child's targetlist doesn't match
MergeAppend");
Although I think, that more accurate check will repeat work done in
prepare_sort_from_pathkeys().
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/
Teodor Sigaev <teodor@sigaev.ru> writes:
SELECT
t1, t2
FROM (
SELECT t1, t2 FROM wow
UNION ALL
SELECT 'a', 'a' FROM wow
) t
ORDER BY t1, t2;
Hmm, interesting.
It's seems to me that check in create_merge_append_plan() is too restrictive:
if (memcmp(sortColIdx, node->sortColIdx,
numsortkeys * sizeof(AttrNumber)) != 0)
elog(ERROR, "MergeAppend child's targetlist doesn't match
MergeAppend");
No, it isn't. That code is fine; the problem is that
add_child_rel_equivalences is generating an invalid state of the
EquivalenceClass structures by adding equal items to two different
EquivalenceClasses. We need to rethink what that routine is doing.
It's definitely wrong for it to add constant items; here, that would
imply injecting t1 = 'a' and t2 = 'a' conditions, which is not correct.
And the same problem is observed if second
select is replaced by "SELECT t1, t1 FROM wow".
And this one is a bit nasty too, since it would still add equal items
to two different ECs, leading to the conclusion that they should be
merged, ie t1 = t2, which is likewise wrong.
Not immediately sure what to do about this. The quick-and-dirty fix
would be to only apply add_child_rel_equivalences to simple inheritance
child relations, for which the added items must be Vars and must be
different (which is what that code is expecting). Seems like a bit of a
shame to lobotomize planning for UNION cases, though. Maybe we need a
more complicated representation of child EquivalenceClass members.
regards, tom lane
I wrote:
Not immediately sure what to do about this. The quick-and-dirty fix
would be to only apply add_child_rel_equivalences to simple inheritance
child relations, for which the added items must be Vars and must be
different (which is what that code is expecting). Seems like a bit of a
shame to lobotomize planning for UNION cases, though. Maybe we need a
more complicated representation of child EquivalenceClass members.
After some thought and experimentation, the best fix seems to be to
eliminate the ambiguity by wrapping subquery outputs in PlaceHolderVars
whenever there is a risk of confusion. The attached crude patch fixes
both test cases. It needs work (more comments and a regression test
case would be good), but barring objection I'll push forward with doing
it this way.
regards, tom lane
After some thought and experimentation, the best fix seems to be to
eliminate the ambiguity by wrapping subquery outputs in PlaceHolderVars
whenever there is a risk of confusion. The attached crude patch fixes
both test cases. It needs work (more comments and a regression test
case would be good), but barring objection I'll push forward with doing
it this way.
Thank you, your patch fixes original query too.
--
Teodor Sigaev E-mail: teodor@sigaev.ru
WWW: http://www.sigaev.ru/