P20-831: Optimize user search by username

[artem-vavilov]

Summary

Testing the users table locally with over 3.25 million records

Before (9483.1ms)

EXPLAIN for: SELECT `users`.* FROM `users` WHERE `users`.`username` LIKE '%art%'
+----+-------------+-------+------------+------+---------------+------+---------+------+---------+----------+-------------+
| id | select_type | table | partitions | type | possible_keys | key  | key_len | ref  | rows    | filtered | Extra       |
+----+-------------+-------+------------+------+---------------+------+---------+------+---------+----------+-------------+
|  1 | SIMPLE      | users | NULL       | ALL  | NULL          | NULL | NULL    | NULL | 3135067 |    11.11 | Using where |
+----+-------------+-------+------------+------+---------------+------+---------+------+---------+----------+-------------+

After (0.6ms)

EXPLAIN for: SELECT `users`.* FROM `users` WHERE `users`.`username` = 'artem'
+----+-------------+-------+------------+------+----------------------------------------+----------------------------------------+---------+-------+------+----------+-------+
| id | select_type | table | partitions | type | possible_keys                          | key                                    | key_len | ref   | rows | filtered | Extra |
+----+-------------+-------+------------+------+----------------------------------------+----------------------------------------+---------+-------+------+----------+-------+
|  1 | SIMPLE      | users | NULL       | ref  | index_users_on_username_and_deleted_at | index_users_on_username_and_deleted_at | 768     | const |    1 |    100.0 | NULL  |
+----+-------------+-------+------------+------+----------------------------------------+----------------------------------------+---------+-------+------+----------+-------+

Links

• JIRA ticket: P20-831

[carl-codeorg]

What would the performance look like if we got rid of the wildcards entirely and did User.arel_table[:username].eq(params[:usernameFilter])? In your test, the improvement was pretty huge by removing the leading %. If we get another large improvement by going to .eq it might be worth switching over, considering the large size of the users table and the fact that the CX team says they normally have the exact username they want to search.

[carl-codeorg]

LGTM! Thanks for taking this on!

[carl-codeorg]

Btw after this is deployed, it's probably worth posting in #customer-support to inform them of the change.