I was curious. And as we all know, curiosity has a reputation for killing cats.
So, which is the fastest way to skin a cat?
The cat-skinning environment for this test:
- PostgreSQL 9.0 on Debian Squeeze with decent RAM and settings.
- 6.000 students, 24.000 club memberships (data copied from a similar database with real life data.)
- Slight diversion from the naming schema in the question:
student.idisstudent.stud_idandclub.idisclub.club_idhere. - I named the queries after their author in this thread.
- I ran all queries a couple of times to populate the cache, then I picked the best of 5 with
EXPLAIN ANALYZE. - Relevant indexes (should be the optimum – as long as we lack fore-knowledge which clubs will be queried):
ALTER TABLE student ADD CONSTRAINT student_pkey PRIMARY KEY(stud_id );
ALTER TABLE student_club ADD CONSTRAINT sc_pkey PRIMARY KEY(stud_id, club_id);
ALTER TABLE club ADD CONSTRAINT club_pkey PRIMARY KEY(club_id );
CREATE INDEX sc_club_id_idx ON student_club (club_id);
club_pkey is not required by most queries here.
Primary keys implement unique indexes automatically In PostgreSQL.
The last index is to make up for this known shortcoming of multi-column indexes on PostgreSQL:
A multicolumn B-tree index can be used with query conditions that
involve any subset of the index’s columns, but the index is most
efficient when there are constraints on the leading (leftmost) columns.
Results
Total runtimes from EXPLAIN ANALYZE.
1) Martin 2: 44.594 ms
SELECT s.stud_id, s.name
FROM student s
JOIN student_club sc USING (stud_id)
WHERE sc.club_id IN (30, 50)
GROUP BY 1,2
HAVING COUNT(*) > 1;
2) Erwin 1: 33.217 ms
SELECT s.stud_id, s.name
FROM student s
JOIN (
SELECT stud_id
FROM student_club
WHERE club_id IN (30, 50)
GROUP BY 1
HAVING COUNT(*) > 1
) sc USING (stud_id);
3) Martin 1: 31.735 ms
SELECT s.stud_id, s.name
FROM student s
WHERE student_id IN (
SELECT student_id
FROM student_club
WHERE club_id = 30
INTERSECT
SELECT stud_id
FROM student_club
WHERE club_id = 50
);
4) Derek: 2.287 ms
SELECT s.stud_id, s.name
FROM student s
WHERE s.stud_id IN (SELECT stud_id FROM student_club WHERE club_id = 30)
AND s.stud_id IN (SELECT stud_id FROM student_club WHERE club_id = 50);
5) Erwin 2: 2.181 ms
SELECT s.stud_id, s.name
FROM student s
WHERE EXISTS (SELECT * FROM student_club
WHERE stud_id = s.stud_id AND club_id = 30)
AND EXISTS (SELECT * FROM student_club
WHERE stud_id = s.stud_id AND club_id = 50);
6) Sean: 2.043 ms
SELECT s.stud_id, s.name
FROM student s
JOIN student_club x ON s.stud_id = x.stud_id
JOIN student_club y ON s.stud_id = y.stud_id
WHERE x.club_id = 30
AND y.club_id = 50;
The last three perform pretty much the same. 4) and 5) result in the same query plan.
Late Additions
Fancy SQL, but the performance can’t keep up:
7) ypercube 1: 148.649 ms
SELECT s.stud_id, s.name
FROM student AS s
WHERE NOT EXISTS (
SELECT *
FROM club AS c
WHERE c.club_id IN (30, 50)
AND NOT EXISTS (
SELECT *
FROM student_club AS sc
WHERE sc.stud_id = s.stud_id
AND sc.club_id = c.club_id
)
);
8) ypercube 2: 147.497 ms
SELECT s.stud_id, s.name
FROM student AS s
WHERE NOT EXISTS (
SELECT *
FROM (
SELECT 30 AS club_id
UNION ALL
SELECT 50
) AS c
WHERE NOT EXISTS (
SELECT *
FROM student_club AS sc
WHERE sc.stud_id = s.stud_id
AND sc.club_id = c.club_id
)
);
As expected, those two perform almost the same. Query plan results in table scans, the planner doesn’t find a way to use the indexes here.
9) wildplasser 1: 49.849 ms
WITH RECURSIVE two AS (
SELECT 1::int AS level
, stud_id
FROM student_club sc1
WHERE sc1.club_id = 30
UNION
SELECT two.level + 1 AS level
, sc2.stud_id
FROM student_club sc2
JOIN two USING (stud_id)
WHERE sc2.club_id = 50
AND two.level = 1
)
SELECT s.stud_id, s.student
FROM student s
JOIN two USING (studid)
WHERE two.level > 1;
Fancy SQL, decent performance for a CTE. Very exotic query plan.
10) wildplasser 2: 36.986 ms
WITH sc AS (
SELECT stud_id
FROM student_club
WHERE club_id IN (30,50)
GROUP BY stud_id
HAVING COUNT(*) > 1
)
SELECT s.*
FROM student s
JOIN sc USING (stud_id);
CTE variant of query 2). Surprisingly, it can result in a slightly different query plan with the exact same data. I found a sequential scan on student, where the subquery-variant used the index.
11) ypercube 3: 101.482 ms
Another late addition ypercube. It is positively amazing, how many ways there are.
SELECT s.stud_id, s.student
FROM student s
JOIN student_club sc USING (stud_id)
WHERE sc.club_id = 10 -- member in 1st club ...
AND NOT EXISTS (
SELECT *
FROM (SELECT 14 AS club_id) AS c -- can't be excluded for missing the 2nd
WHERE NOT EXISTS (
SELECT *
FROM student_club AS d
WHERE d.stud_id = sc.stud_id
AND d.club_id = c.club_id
)
);
12) erwin 3: 2.377 ms
ypercube’s 11) is actually just the mind-twisting reverse approach of this simpler variant, that was also still missing. Performs almost as fast as the top cats.
SELECT s.*
FROM student s
JOIN student_club x USING (stud_id)
WHERE sc.club_id = 10 -- member in 1st club ...
AND EXISTS ( -- ... and membership in 2nd exists
SELECT *
FROM student_club AS y
WHERE y.stud_id = s.stud_id
AND y.club_id = 14
);
13) erwin 4: 2.375 ms
Hard to believe, but here’s another, genuinely new variant. I see potential for more than two memberships, but it also ranks among the top cats with just two.
SELECT s.*
FROM student AS s
WHERE EXISTS (
SELECT *
FROM student_club AS x
JOIN student_club AS y USING (stud_id)
WHERE x.stud_id = s.stud_id
AND x.club_id = 14
AND y.club_id = 10
);
Dynamic number of club memberships
In other words: varying number of filters. This question asked for exactly two club memberships. But many use cases have to prepare for a varying number. See: