[PATCH] Restrict partial unique index estimates to base quals

From: Chengpeng Yan <chengpeng_yan(at)outlook(dot)com>
To: "pgsql-hackers(at)lists(dot)postgresql(dot)org" <pgsql-hackers(at)lists(dot)postgresql(dot)org>
Cc: David Rowley <dgrowleyml(at)gmail(dot)com>
Subject: [PATCH] Restrict partial unique index estimates to base quals
Date: 2026-07-26 08:01:30
Message-ID: EC9A6C8C-C671-45F2-A580-0276566CE677@outlook.com
Views: Whole Thread | Raw Message | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

Hi,

Attached is v1 of a patch fixing selectivity estimates from partial
unique indexes. Currently, a partial index key can be treated as unique
when predOK was established using join clauses. For example:

CREATE TABLE parent (id int);
CREATE TABLE child (parent_id int, code int);

-- Code is unique only when parent_id is not NULL.
CREATE UNIQUE INDEX ON child (code) WHERE parent_id IS NOT NULL;

INSERT INTO parent VALUES (1);
-- Insert 1000 rows outside the predicate, and one inside it.
INSERT INTO child SELECT NULL, 42 FROM generate_series(1, 1000);
INSERT INTO child VALUES (1, 42);
ANALYZE parent, child;

SET enable_indexscan = off;
SET enable_bitmapscan = off;

EXPLAIN
SELECT * FROM parent p JOIN child c ON c.parent_id = p.id
WHERE c.code = 42;

An unpatched server estimates one row for the child scan, although code
= 42 matches all 1001 rows; the join eliminates the 1000 NULL rows only
later.

Bug #17975 exposed the same proof-scope problem in relation uniqueness
[1]. Commit f6345f03f35 fixed it by rejecting all partial indexes in
relation_has_unique_index_for(). That was a simple and safe fix for a
correctness bug that needed backpatching.

We could do the same here, but base-proven uniqueness is still valid and
useful for estimates. The patch therefore adds predOKBase, proved from
baserestrictinfo only, rather than rejecting all partial indexes.
Simple-column and expression-index estimates use predOKBase. predOK
retains exactly the same result: a base-only proof sets both flags;
otherwise, the existing join-aware proof is still used. This is the
same base-only distinction David proposed for relation uniqueness [2].
The child scan above is then estimated at 1001 rows.

That relation-uniqueness change is not part of v1, but I would be happy
to follow up if useful.

[1] https://postgr.es/m/17975-98a90c156f25c952%40postgresql.org
[2] https://postgr.es/m/CAApHDvo4s_1TzzYzd5VKgfrnPpOm18HBT544L%3DxTv0jcVrqFXg%40mail.gmail.com

--
Best regards,
Chengpeng Yan

Attachment Content-Type Size
v1-0001-Restrict-partial-unique-index-estimates-to-base-q.patch application/octet-stream 13.8 KB

Browse pgsql-hackers by date

  From Date Subject
Previous Message Mats Kindahl 2026-07-26 05:47:28 Recovery does not honor io_combine_limit, causing IOPS saturation