The following bug has been logged on the website:
Bug reference: 19751
Logged by: N J
Email address: 1482694023@qq.com
PostgreSQL version: 18.6
Operating system: Windows 11 64-bit
Description:
Environment:
PostgreSQL: [fill in your exact version]
OS: Windows 11 64-bit
Steps to reproduce:
1. Create test table and insert data:
CREATE TEMP TABLE readings (
region_id integer NOT NULL,
reading integer NOT NULL
);
INSERT INTO readings (region_id, reading) VALUES
(2, 7),
(2, 1);
2. Create index and collect statistics:
CREATE INDEX ON readings (region_id, reading);
ANALYZE readings;
3. Force index scan and execute the query:
SET enable_seqscan = off;
SELECT MIN(reading) AS minimum
FROM (
SELECT reading
FROM readings
WHERE region_id = 2
) AS scoped
ORDER BY minimum
FETCH FIRST 1 ROW WITH TIES;
Observed result:
ERROR: more than one row returned by a subquery used as an expression
SQL state: 21000
Expected result:
minimum
---------
1
(1 row)
Bug analysis:
This is a valid standard SQL query. The subquery is only used as a derived
table in the FROM clause, there is no scalar subquery in the statement. The
21000 cardinality error should never be raised here.
The error only triggers under these conditions:
- Sequential scan is disabled, so the optimizer uses an index scan to
optimize the MIN() aggregate
- The query includes FETCH FIRST 1 ROW WITH TIES
- The inner derived table returns multiple rows
It appears that when the optimizer combines the index-optimized MIN() path
with the WITH TIES logic, it incorrectly generates an execution plan that
treats the internal index scan subplan as a scalar subquery, resulting in a
spurious error when the subplan returns multiple rows. This is an optimizer
correctness bug.
