When using partitioning in PostgreSQL, where master is the master partition table
CREATE TABLE master
(
_id numeric,
name character varying
);
and having two sub tables
partition_1
CREATE TABLE partition_1
(
-- _id numeric,
-- name character varying,
CONSTRAINT partition_1_check CHECK (_id < 1000)
)
INHERITS (master);
CREATE INDEX partition_1_id_idx
ON partition_1
USING btree
(_id);
partition_2
CREATE TABLE partition_2
(
-- _id numeric,
-- name character varying,
CONSTRAINT partition_2_check CHECK (_id >= 1000)
)
INHERITS (master);
CREATE INDEX partition_2_id_idx
ON partition_2
USING btree
(_id);
and a table for some ids (1,3) in this example
CREATE TABLE _ids
(
_id numeric NOT NULL,
CONSTRAINT ids_pkey PRIMARY KEY (_id)
)
the statement
EXPLAIN SELECT * FROM master WHERE _id IN (SELECT * FROM _ids)
yields a seq scan of both partitions regardless whether _ids contains elements from partition_1/2 or not.
Hash Semi Join (cost=39.48..141.14 rows=2621 width=14)
Hash Cond: (master._id = _ids._id)
-> Append (cost=0.00..62.00 rows=4001 width=14)
-> Seq Scan on master (cost=0.00..0.00 rows=1 width=14)
-> Seq Scan on partition_1 (cost=0.00..30.98 rows=1998 width=13)
-> Seq Scan on partition_2 (cost=0.00..31.02 rows=2002 width=15)
-> Hash (cost=23.10..23.10 rows=1310 width=32)
-> Seq Scan on _ids (cost=0.00..23.10 rows=1310 width=32)
If I instead use something like
SELECT * FROM master WHERE _id IN (1,3)
I get the desired result:
Append (cost=0.00..17.40 rows=5 width=13)
-> Seq Scan on master (cost=0.00..0.00 rows=1 width=14)
Filter: (_id = ANY ('{1,3}'::numeric[]))
-> Bitmap Heap Scan on partition_1 (cost=8.59..17.40 rows=4 width=13)
Recheck Cond: (_id = ANY ('{1,3}'::numeric[]))
-> Bitmap Index Scan on partition_1_id_idx (cost=0.00..8.58 rows=4 width=0)
Index Cond: (_id = ANY ('{1,3}'::numeric[]))
How do I get PostgreSQL to use constraint exclusion correctly ?
Note: I have constraint_exclusion set to partition
The partitions are chosen when Postgres compiles that query. When you use constants, then the engine knows where the data lies. When you use a subquery, then Postgres doesn't know.
Hence, using the subquery prevents the engine from taking advantage of the partitions.