[Bug] Wrong results: GPORCA evaluates a window function in a correlated aggregate subquery over all groups
Nessuno ha ancora preso questa issue.
Valutazione
- Difficoltà
- 4/5
- Tempo stimato
- 3-5 giorni
- Idoneità per principianti
- 40/100
Direzione di ricerca
The issue is in GPORCA's decorrelation logic for subqueries containing window functions. Start by examining the query planner code that transforms correlated subqueries, particularly around GroupAggregate and WindowAgg nodes. Look at the test case in the issue to understand the wrong grouping. The fix likely involves ensuring the window function is evaluated after aggregation, not across groups. Run the provided SQL to reproduce the bug and verify with optimizer on/off.
Scritto dal modello di indicizzazione a partire dal testo della issue.
Descrizione
Apache Cloudberry version
main; REL_2_STABLE
What happened
With GPORCA (optimizer = on), a correlated scalar subquery that combines an aggregate with a window function returns a wrong value. The subquery has no GROUP BY, so it produces exactly one row per outer row, and count(*) over () inside it must be 1. GPORCA decorrelates the subquery into a GroupAggregate grouped by the correlation column and puts the WindowAgg on top. The window then runs over all groups instead of the single aggregate row.
This gives a wrong value in the target list and wrong rows in WHERE. The Postgres planner (optimizer = off) computes the target-list form correctly (see below for WHERE).
What you think should happen instead
count(*) over () inside the subquery should see only the single result row of the ungrouped aggregate, as it does with the Postgres planner.
How to reproduce
create table w1(a int, b int); insert into w1 values (4,1);
create table w2(a int); insert into w2 values (1),(1),(2);
analyze w1; analyze w2;
set optimizer = on;
select a, b, (select sum(w2.a) + count(*) over () from w2 where w2.a = w1.b) from w1;
-- a | b | ?column?
-- ---+---+----------
-- 4 | 1 | 4 <-- WRONG, expected 3 (sum = 2, count(*) over () = 1)
select a, b from w1 where w1.a > (select sum(w2.a) + count(*) over () from w2 where w2.a = w1.b);
-- (0 rows) <-- WRONG, expected 4 | 1
With optimizer = off, the target-list query returns the expected 4 | 1 | 3. The WHERE query is correct on main only when the pull-up is guarded against window functions (with patch from #1933). On REL_2_STABLE, the Postgres planner also returns 0 rows for it.
Plan with optimizer = on:
explain (costs off)
select a, b from w1 where w1.a > (select sum(w2.a) + count(*) over () from w2 where w2.a = w1.b);
Hash Join
Join Filter: (w1.a > ((sum(w2.a)) + count(*) OVER (?)))
-> WindowAgg
-> GroupAggregate
Group Key: w2.a <-- window is computed across all groups
SET optimizer = off; fixes the target-list form. On REL_2_STABLE it does not fix the WHERE form.
Operating System
any
Anything else
No response
Are you willing to submit PR?
- Yes, I am willing to submit a PR!
Code of Conduct
- I agree to follow this project's Code of Conduct.
- Lingua principale
- C
- Stelle
- 1.4k
- Fork
- 248
- Merge medio
- 4g 10h
- PR unite (30g)
- 40
Guida per i contributori
Apri la guida per i contributori
Come iniziare
- Leggi tutta la issue e poi la guida ai contributi del progetto.
- Commenta sulla issue per dire che te ne occupi tu — evita che due persone facciano lo stesso lavoro.
- Fai un fork del repository e lavora su un branch.
- Apri una pull request che faccia riferimento al numero della issue.
Altre issue di apache/cloudberry
-
type: Bug
Difficoltà 2/5 1-3 ore Idoneità per principianti 76/100
apache/cloudberry#1885 · 2 reazioni ·
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 86/100
apache/cloudberry#1825 ·
-
type: Bug
Difficoltà 3/5 1-2 giorni Idoneità per principianti 65/100
apache/cloudberry#2048 · 1 reazione ·
-
type: Bug
Difficoltà 4/5 3-5 giorni Idoneità per principianti 45/100
apache/cloudberry#2046 · 1 commento ·
-
type: Bug
Difficoltà 4/5 3-5 giorni Idoneità per principianti 52/100
apache/cloudberry#2026 · 1 commento · 1 reazione ·
Tutte le issue di apache/cloudberry
Issue simili
-
bug
Difficoltà 2/5 1-3 ore Idoneità per principianti 75/100
bradcypert/plum#53 ·
-
Component: GLib
Difficoltà 2/5 1-3 ore Idoneità per principianti 70/100
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 75/100
-
Status: Opened
Difficoltà 2/5 1-3 ore Idoneità per principianti 70/100
-
Difficoltà 2/5 1-3 ore Idoneità per principianti 75/100
nextbsd/nextbsd-userland#285 ·