Regression in expressions refactor causes database queries to run COUNT('*') instead of COUNT(*)

[Adam Johnson]

Performance regression on 1.8 expression refactor. Count stopped outputting COUNT(*) and started outputting COUNT('*') due to using a Value. Possibly unnoticed until now because it doesn't affect the most popular database options much - I can't measure a performance drop for this on MySQL / InnoDb.

However I have managed to measure a performance drop on MariaDB + Aria (=MySQL fork + MyISAM fork):

adamj@localhost [3]> select count(*) from count_test;
+----------+
| count(*) |
+----------+
|    10000 |
+----------+
1 row in set (0.00 sec)

adamj@localhost [10]> select benchmark(100 * 1000 * 1000, (select count(*) from count_test));
+-----------------------------------------------------------------+
| benchmark(100 * 1000 * 1000, (select count(*) from count_test)) |
+-----------------------------------------------------------------+
|                                                               0 |
+-----------------------------------------------------------------+
1 row in set (0.86 sec)

adamj@localhost [11]> select benchmark(100 * 1000 * 1000, (select count('*') from count_test));
+-------------------------------------------------------------------+
| benchmark(100 * 1000 * 1000, (select count('*') from count_test)) |
+-------------------------------------------------------------------+
|                                                                 0 |
+-------------------------------------------------------------------+
1 row in set (1.23 sec)

This is because MyISAM / Aria store the count in a metadata variable but once you introduce the expression it figures it has to do a table scan. The situation would of course only get worse with more rows, 10000 is tiny.

[Adam Johnson]

Pull request: ​https://github.com/django/django/pull/5260

[Adam Johnson]

I'm not 100% confident in my performance measurement there, I just tried on a multi-million row table on my production DB converted to MyISAM and COUNT(*) measured the same as COUNT('*'). However since there are so many database versions and storage engines out there I think it's probable this caused a regression somewhere.

[Tim Graham <timograham@…>]

In 3fe3887a:

Fixed #25377 -- Changed Count queries to execute COUNT(*) instead of COUNT('*').

[Tim Graham <timograham@…>]

In 3c2c74f:

[1.8.x] Fixed #25377 -- Changed Count queries to execute COUNT(*) instead of COUNT('*').

Backport of 3fe3887a2ed94f7b15be769f6d81571031ec5627 from master