GROUP BY and COUNT in PostgreSQL

Question

The query:

SELECT COUNT(*) as count_all, 
       posts.id as post_id 
FROM posts 
  INNER JOIN votes ON votes.post_id = posts.id 
GROUP BY posts.id;

Returns n records in Postgresql:

 count_all | post_id
-----------+---------
 1         | 6
 3         | 4
 3         | 5
 3         | 1
 1         | 9
 1         | 10
(6 rows)

I just want to retrieve the number of records returned: 6.

I used a subquery to achieve what I want, but this doesn't seem optimum:

SELECT COUNT(*) FROM (
    SELECT COUNT(*) as count_all, posts.id as post_id 
    FROM posts 
    INNER JOIN votes ON votes.post_id = posts.id 
    GROUP BY posts.id
) as x;

How would I get the number of records in this context right in PostgreSQL?

Answer 1

I think you just need COUNT(DISTINCT post_id) FROM votes.

See "4.2.7. Aggregate Expressions" section in http://www.postgresql.org/docs/current/static/sql-expressions.html.

EDIT: Corrected my careless mistake per Erwin's comment.

Answer 2

There is also EXISTS:

SELECT count(*) AS post_ct
FROM   posts p
WHERE  EXISTS (SELECT FROM votes v WHERE v.post_id = p.id);

In Postgres and with multiple entries on the n-side like you probably have, it's generally faster than count(DISTINCT post_id):

SELECT count(DISTINCT p.id) AS post_ct
FROM   posts p
JOIN   votes v ON v.post_id = p.id;

The more rows per post there are in votes, the bigger the difference in performance. Test with EXPLAIN ANALYZE.

count(DISTINCT post_id) has to read all rows, sort or hash them, and then only consider the first per identical set. EXISTS will only scan votes (or, preferably, an index on post_id) until the first match is found.

If every post_id in votes is guaranteed to be present in the table posts (referential integrity enforced with a foreign key constraint), this short form is equivalent to the longer form:

SELECT count(DISTINCT post_id) AS post_ct
FROM   votes;

May actually be faster than the EXISTS query with no or few entries per post.

The query you had works in simpler form, too:

SELECT count(*) AS post_ct
FROM  (
    SELECT FROM posts 
    JOIN   votes ON votes.post_id = posts.id 
    GROUP  BY posts.id
    ) sub;

Benchmark

To verify my claims I ran a benchmark on my test server with limited resources. All in a separate schema:

Test setup

Fake a typical post / vote situation:

CREATE SCHEMA y;
SET search_path = y;

CREATE TABLE posts (
  id   int PRIMARY KEY
, post text
);

INSERT INTO posts
SELECT g, repeat(chr(g%100 + 32), (random()* 500)::int)  -- random text
FROM   generate_series(1,10000) g;

DELETE FROM posts WHERE random() > 0.9;  -- create ~ 10 % dead tuples

CREATE TABLE votes (
  vote_id serial PRIMARY KEY
, post_id int REFERENCES posts(id)
, up_down bool
);

INSERT INTO votes (post_id, up_down)
SELECT g.* 
FROM  (
   SELECT ((random()* 21)^3)::int + 1111 AS post_id  -- uneven distribution
        , random()::int::bool AS up_down
   FROM   generate_series(1,70000)
   ) g
JOIN   posts p ON p.id = g.post_id;

All of the following queries returned the same result (8093 of 9107 posts had votes).
I ran 4 tests with EXPLAIN ANALYZE ant took the best of five on Postgres 9.1.4 with each of the three queries and appended the resulting total runtimes.

1. As is.

2. After ..

   ANALYZE posts;
   ANALYZE votes;
   

3. After ..

   CREATE INDEX foo on votes(post_id);
   

4. After ..

   VACUUM FULL ANALYZE posts;
   CLUSTER votes using foo;
   

count(*) ... WHERE EXISTS

1. 253 ms
2. 220 ms
3. 85 ms -- winner (seq scan on posts, index scan on votes, nested loop)
4. 85 ms

count(DISTINCT x) - long form with join

1. 354 ms
2. 358 ms
3. 373 ms -- (index scan on posts, index scan on votes, merge join)
4. 330 ms

count(DISTINCT x) - short form without join

1. 164 ms
2. 164 ms
3. 164 ms -- (always seq scan)
4. 142 ms

Best time for original query in question:

• 353 ms

For simplified version:

• 348 ms

@wildplasser's query with a CTE uses the same plan as the long form (index scan on posts, index scan on votes, merge join) plus a little overhead for the CTE. Best time:

• 366 ms

Index-only scans in the upcoming PostgreSQL 9.2 can improve the result for each of these queries, most of all for EXISTS.

Related, more detailed benchmark for Postgres 9.5 (actually retrieving distinct rows, not just counting):

• Select first row in each GROUP BY group?

Answer 3

Using OVER() and LIMIT 1:

SELECT COUNT(1) OVER()
FROM posts 
   INNER JOIN votes ON votes.post_id = posts.id 
GROUP BY posts.id
LIMIT 1;

Answer 4

WITH uniq AS (
        SELECT DISTINCT posts.id as post_id
        FROM posts
        JOIN votes ON votes.post_id = posts.id
        -- GROUP BY not needed anymore
        -- GROUP BY posts.id
        )
SELECT COUNT(*)
FROM uniq;

Answer 5

For followers, I like the OP's inner query method:

SELECT COUNT(*) FROM (
    SELECT COUNT(*) as count_all, posts.id as post_id 
    FROM posts 
    INNER JOIN votes ON votes.post_id = posts.id 
    GROUP BY posts.id
) as x;

Since then you can use HAVING in there as well:

SELECT COUNT(*) FROM (
    SELECT COUNT(*) as count_all, posts.id as post_id 
    FROM posts 
    INNER JOIN votes ON votes.post_id = posts.id 
    GROUP BY posts.id HAVING count(*) > 1
) as x;

Or the equivalent CTE

with posts_coalesced as (
     SELECT COUNT(*) as count_all, posts.id as post_id 
        FROM posts 
        INNER JOIN votes ON votes.post_id = posts.id 
        GROUP BY posts.id )

select count(*) from posts_coalesced;