Tim Biegeleisen Tim Biegeleisen - 5 months ago 9
SQL Question

Performance implications of allowing alias to be used in HAVING clause

I made a bit of a fool out of myself earlier today on this question. The question was using SQL Server, and the correct answer involved adding a

HAVING
clause. The initial mistake I made was to think that an alias in the
SELECT
statement could be used in the
HAVING
clause, which is not allowed in SQL Server. I made this error because I assumed that SQL Server had the same rules as MySQL, which does allow an alias to be used in the
HAVING
clause.

This got me curious, and I poked around on Stack Overflow and elsewhere, finding a bunch of material explaining why these rules are enforced on the two respective RDBMS. But nowhere did I find an explanation of what the performance implications would be of allowing/disallowing an alias in the
HAVING
clause.

To give a concrete example, I will duplicate the query which occurred in the above-mentioned question:

SELECT students.camID, campus.camName, COUNT(students.stuID) as studentCount
FROM students
JOIN campus
ON campus.camID = students.camID
GROUP BY students.camID, campus.camName
HAVING COUNT(students.stuID) > 3
ORDER BY studentCount


What would be the performance implications of using an alias in the
HAVING
clause instead of re specifying the
COUNT
? This question can answered directly in MySQL, and hopefully someone could give insight into what would happen in SQL if it were to support the alias in the
HAVING
clause.

This is a rare instance where it might be OK to tag a SQL question with both MySQL and SQL Server, so enjoy this moment in the sun.

Answer

Narrowly focused on just that particular query, and with sample data loaded below. This does address some other queries such as the count(distinct ...) mentioned by others.

The alias in the HAVING appears to either slightly outperform or quite a bit outperform its alternative (depending on the query).

This uses a pre-existing table with about 5 million rows in it created quickly via this answer of mine which takes 3 to 5 minutes.

Resulting structure:

CREATE TABLE `ratings` (
  `id` int(11) NOT NULL AUTO_INCREMENT,
  `thing` int(11) NOT NULL,
  PRIMARY KEY (`id`)
) ENGINE=InnoDB AUTO_INCREMENT=5046214 DEFAULT CHARSET=utf8;

But using INNODB instead. Creates the expected INNODB gap anomaly due to the range reservation inserts. Just saying, but makes no difference. 4.7 Million rows.

Modify the table to get near Tim's assumed schema.

rename table ratings to students; -- not exactly instanteous (a COPY)
alter table students add column camId int; -- get it near Tim's schema
-- don't add the `camId` index yet

The following will take a while. Run it again and again in chunks or else your connection may timeout. The timeout is due to 5 million rows without a LIMIT clause in the update statement. Note, we do have a LIMIT clause.

So we are doing it in half a million row iterations. Sets a column to a random number between 1 and 20

update students set camId=floor(rand()*20+1) where camId is null limit 500000; -- well that took a while (no surprise)

Keep running the above until no camId is null.

I ran it like 10 times (the whole thing takes 7 to 10 minutes)

select camId,count(*) from students
group by camId order by 1 ;

1   235641
2   236060
3   236249
4   235736
5   236333
6   235540
7   235870
8   236815
9   235950
10  235594
11  236504
12  236483
13  235656
14  236264
15  236050
16  236176
17  236097
18  235239
19  235556
20  234779

select count(*) from students;
-- 4.7 Million rows

Create a useful index (after the inserts of course).

create index `ix_stu_cam` on students(camId); -- takes 45 seconds

ANALYZE TABLE students; -- update the stats: http://dev.mysql.com/doc/refman/5.7/en/analyze-table.html
-- the above is fine, takes 1 second

Create the campus table.

create table campus
(   camID int auto_increment primary key,
    camName varchar(100) not null
);
insert campus(camName) values
('one'),('2'),('3'),('4'),('5'),
('6'),('7'),('8'),('9'),('ten'),
('etc'),('etc'),('etc'),('etc'),('etc'),
('etc'),('etc'),('etc'),('etc'),('twenty');
-- ok 20 of them

Run the two queries:

SELECT students.camID, campus.camName, COUNT(students.id) as studentCount 
FROM students 
JOIN campus 
    ON campus.camID = students.camID 
GROUP BY students.camID, campus.camName 
HAVING COUNT(students.id) > 3 
ORDER BY studentCount; 
-- run it many many times, back to back, 5.50 seconds, 20 rows of output

and

SELECT students.camID, campus.camName, COUNT(students.id) as studentCount 
FROM students 
JOIN campus 
    ON campus.camID = students.camID 
GROUP BY students.camID, campus.camName 
HAVING studentCount > 3 
ORDER BY studentCount; 
-- run it many many times, back to back, 5.50 seconds, 20 rows of output

So the times are identical. Ran each a dozen times.

The EXPLAIN output is the same for both

+----+-------------+----------+------+---------------+------------+---------+----------------------+--------+---------------------------------+
| id | select_type | table    | type | possible_keys | key        | key_len | ref                  | rows   | Extra                           |
+----+-------------+----------+------+---------------+------------+---------+----------------------+--------+---------------------------------+
|  1 | SIMPLE      | campus   | ALL  | PRIMARY       | NULL       | NULL    | NULL                 |     20 | Using temporary; Using filesort |
|  1 | SIMPLE      | students | ref  | ix_stu_cam    | ix_stu_cam | 5       | bigtest.campus.camID | 123766 | Using index                     |
+----+-------------+----------+------+---------------+------------+---------+----------------------+--------+---------------------------------+

Using the AVG() function, I am getting about a 12% increase in performance with the alias in the having (with identical EXPLAIN output) from the following two queries.

SELECT students.camID, campus.camName, avg(students.id) as studentAvg 
FROM students 
JOIN campus 
    ON campus.camID = students.camID 
GROUP BY students.camID, campus.camName 
HAVING avg(students.id) > 2200000 
ORDER BY students.camID; 
-- avg time 7.5

explain 

SELECT students.camID, campus.camName, avg(students.id) as studentAvg 
FROM students 
JOIN campus 
    ON campus.camID = students.camID 
GROUP BY students.camID, campus.camName 
HAVING studentAvg > 2200000
ORDER BY students.camID;
-- avg time 6.5

And lastly, the DISTINCT:

SELECT students.camID, count(distinct students.id) as studentDistinct 
FROM students 
JOIN campus 
    ON campus.camID = students.camID 
GROUP BY students.camID 
HAVING count(distinct students.id) > 1000000 
ORDER BY students.camID; -- 10.6   10.84   12.1   11.49   10.1   9.97   10.27   11.53   9.84 9.98
-- 9.9

 SELECT students.camID, count(distinct students.id) as studentDistinct 
 FROM students 
 JOIN campus 
    ON campus.camID = students.camID 
 GROUP BY students.camID 
 HAVING studentDistinct > 1000000 
 ORDER BY students.camID; -- 6.81    6.55   6.75   6.31   7.11 6.36   6.55
-- 6.45

The alias in the having consistently runs 35% faster with the same EXPLAIN output. Seen below. So the same Explain output has been shown twice to not result in the same performance, but as a general clue.

+----+-------------+----------+-------+---------------+------------+---------+----------------------+--------+----------------------------------------------+
| id | select_type | table    | type  | possible_keys | key        | key_len | ref                  | rows   | Extra                                        |
+----+-------------+----------+-------+---------------+------------+---------+----------------------+--------+----------------------------------------------+
|  1 | SIMPLE      | campus   | index | PRIMARY       | PRIMARY    | 4       | NULL                 |     20 | Using index; Using temporary; Using filesort |
|  1 | SIMPLE      | students | ref   | ix_stu_cam    | ix_stu_cam | 5       | bigtest.campus.camID | 123766 | Using index                                  |
+----+-------------+----------+-------+---------------+------------+---------+----------------------+--------+----------------------------------------------+

The Optimizer Appears to favor the alias in the having at the moment, especially for the DISTINCT.

Comments