3

I have a table with several columns, say a, b, c, d that should be searchable. The problem arises when I need to search a, b, c separately from d (and vise-versa). AFAIK, there's no way to achieve this using one composite fulltext index on all columns, so I create two separate indexes like this:

CREATE FULLTEXT INDEX idx1 ON content (a, b, c);
CREATE FULLTEXT INDEX idx2 ON content (d);

Now I can search the first and second one successfully. For both of them, I would use the following command:

SELECT * FROM content 
WHERE MATCH(a, b, c) AGAINST ('keyword')
AND MATCH(d) AGAINST ('keyword');

explain tells me this:

+----+-------------+---------+----------+---------------+------+---------+------+------+-------------+
| id | select_type | table   | type     | possible_keys | key  | key_len | ref  | rows | Extra       |
+----+-------------+---------+----------+---------------+------+---------+------+------+-------------+
|  1 | SIMPLE      | content | fulltext | idx1,idx2     | idx1 | 0       |      |    1 | Using where |
+----+-------------+---------+----------+---------------+------+---------+------+------+-------------+

Great! So it's using two indexes, but returns only rows where keyword is present in both inclusive and I need either one, so I change AND to OR and now explain says:

+----+-------------+---------+------+---------------+------+---------+------+------+-------------+
| id | select_type | table   | type | possible_keys | key  | key_len | ref  | rows | Extra       |
+----+-------------+---------+------+---------------+------+---------+------+------+-------------+
|  1 | SIMPLE      | content | ALL  | NULL          | NULL | NULL    | NULL |  128 | Using where |
+----+-------------+---------+------+---------------+------+---------+------+------+-------------+

What? Suddenly, it's doing a full table scan. Why is this happening? What would be the best way to avoid this?

4

Unfortunately, this is how MySQL Query Optimizer treats FULLTEXT indexes. When a MATCH clause is the only clause in the WHERE, the index will be used. When used in conjunction with AND, the index may easily get overlooked.

I wrote about this behavior before in Mysql fulltext search my.cnf optimization

SUGGESTION : Rewrite the query as the union of two FULLTEXT searches

SELECT * FROM content 
WHERE MATCH(a, b, c) AGAINST ('keyword')
UNION
SELECT * FROM content 
WHERE MATCH(d) AGAINST ('keyword');

GIVE IT A TRY !!!

  • Thanks! Indeed, this gives me the desired result, however when run with explain, it now shows 3 actions instead of 1. Does this mean it actually doing more operations, or it comes down to one lookup internally? – bobo Jul 5 '15 at 20:12
  • Yes it does. It should be 3 operations, 1 resultset from MATCH(a, b, c), 1 resultset from MATCH(d), and a merger of the two resultsets. – RolandoMySQLDBA Jul 5 '15 at 20:22
0

Unclear what you want to do.

If you want to see rows with keyword in both d and somewhere in a,b,c, then your AND is a good way to go.

But if you want keyword to be in any of a,b,c,d, then add a third index

FULLTEXT(a,b,c,d)

and change to

MATCH(a,b,c,d) AGAINST('keyword')

For further discussion please specify whether you are using InnoDB or MyISAM; they work differently.

  • This way I'll have 3 indexes, one of which is completely redundant. Plus, on every INSERT, MySQL would have to rebuild all three of them. Seems like too much extra operations for a simple search feature. Using MyISAM. – bobo Jul 7 '15 at 17:48
  • MySQL does not "rebuild" indexes on each INSERT; it just augments the indexes. This is cheap enough not to be a big issue. I think MyISAM does not need the third FULLTEXT, but InnoDB does, hence my question. – Rick James Jul 7 '15 at 19:05
  • That's nice to know, thanks. It doesn't absolutely need the third FULLTEXT, but in that case it's doing full table scan, which I'm trying to avoid. My original question was why does MySQL treat AND so differently from OR in a single query, when internally it should be implemented the same way. There doesn't seem to be much information online about this behavior, except for the question here, which basically says "OR is not optimized, while AND is" – bobo Jul 7 '15 at 22:39
  • 1
    Yes. AND can often be optimized, often with a different index. OR is rarely optimizable. Often UNION is a trick to optimize it, as Rolando points out. – Rick James Jul 8 '15 at 3:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.