mysql count(*) count(val) count(1)比较

今天上网，看到一种说法如下（仅作记录，不做评论）：

如果表中没有主键，那么count(1)比count(*)快
如果有主键，那么count(主键，联合主键)比count(*)快
如果表中只有一个字段，count(*)最快

再摘一篇类似的英文:http://www.mysqlperformanceblog.com/2007/04/10/count-vs-countcol/

Looking at how people are using COUNT(*) and COUNT(col) it looks like most of them think they are synonyms and just using what they happen to like, while there is substantial difference in performance and even query result.
Lets look at the following series of examples:

PLAIN TEXT
SQL:

CREATE TABLE `fact` (
  `i` int(10) UNSIGNED NOT NULL,
  `val` int(11) DEFAULT NULL,
  `val2` int(10) UNSIGNED NOT NULL,
  KEY `i` (`i`)
) ENGINE=MyISAM DEFAULT CHARSET=latin1


mysql> SELECT count(*) FROM fact;
+----------+
| count(*) |
+----------+
|  7340032 |
+----------+
1 row IN SET (0.00 sec)


mysql> SELECT count(val) FROM fact;
+------------+
| count(val) |
+------------+
|    7216582 |
+------------+
1 row IN SET (1.17 sec)


mysql> SELECT count(val2) FROM fact;
+-------------+
| count(val2) |
+-------------+
|     7340032 |
+-------------+

• 1 row IN SET (0.00 sec)

As this is MYISAM table MySQL has cached number of rows in this table. This is why it is able to instantly answer COUNT(*) and
COUNT(val2) queries, but not COUNT(val). Why ? Because val column is not defined as NOT NULL there can be some NULL values in it and so MySQL have to perform table scan to find out. This is also why result is different for the second query.
So COUNT(*) and COUNT(col) queries not only could have substantial performance performance differences but also ask different question.
MySQL Optimizer does good job in this case doing full table scan only if it is needed because column can be NULL.
Now lets try few more queries:
PLAIN TEXT
SQL:

mysql> SELECT count(*) FROM fact WHERE i<10000;
+----------+
| count(*) |
+----------+
|   733444 |
+----------+
1 row IN SET (0.40 sec)

mysql> EXPLAIN SELECT count(*) FROM fact WHERE i<10000 \G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        TABLE: fact
         type: range
possible_keys: i
          KEY: i
      key_len: 4
          ref: NULL
         rows: 691619
        Extra: USING WHERE; USING INDEX
1 row IN SET (0.00 sec)


mysql> SELECT count(val) FROM fact WHERE i<10000;
+------------+
| count(val) |
+------------+
|     720934 |
+------------+
1 row IN SET (1.29 sec)

mysql> EXPLAIN SELECT count(val) FROM fact WHERE i<10000 \G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        TABLE: fact
         type: range
possible_keys: i
          KEY: i
      key_len: 4
          ref: NULL
         rows: 691619
        Extra: USING WHERE
1 row IN SET (0.00 sec)

mysql> SELECT count(val2) FROM fact WHERE i<10000;
+-------------+
| count(val2) |
+-------------+
|      733444 |
+-------------+
1 row IN SET (1.30 sec)

mysql> EXPLAIN SELECT count(val2) FROM fact WHERE i<10000 \G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        TABLE: fact
         type: range
possible_keys: i
          KEY: i
      key_len: 4
          ref: NULL
         rows: 691619
        Extra: USING WHERE

• 1 row IN SET (0.00 sec)

As you can see even if you have where clause performance for count(*) and count(col) can be significantly different. In fact this example shows just 3 times performance difference because all data fits in memory, for IO bound workloads you frequently can see 10 and even 100 times performance difference in this case.
The thing is count(*) query can use covering index even while count(col) can't. Of course you can extend index to be (i,val) and get query to be index covered again but I would use this workaround only if you can't change the query (ie it is third party application) or in case column name is in the query for reason, and you really need count of non-NULL values.
It is worth to note in this case MySQL Optimizer does not do too good job optimizing the query. One could notice (val2) column is not null so count(val2) is same as count(*) and so the query could be run as index covered query. It does not and both queries have to perform row reads in this case.
PLAIN TEXT
SQL:

mysql> ALTER TABLE fact DROP KEY i, ADD KEY(i,val);
Query OK, 7340032 rows affected (37.15 sec)
Records: 7340032  Duplicates: 0  Warnings: 0


mysql> SELECT count(val) FROM fact WHERE i<10000;
+------------+
| count(val) |
+------------+
|     720934 |
+------------+

• 1 row IN SET (0.78 sec)

As you can see extending index helps in this case but it makes query about 2 times slower compared to count(*) one. This is probably because index becomes about two times longer in this case.