I have a USER table with millions of rows. I am implementing a search function that allows someone to look for a user by typing in a username. This autocomplete feature needs to be blazingly fast. Given that, in MySQL, column indexes speed up queries using LIKE {string}%, is the following approach performant enough to return within 200ms? (Note: Memory overhead is not an issue here, username are maximum 30 characters).
Create a USERSEARCH table that has a foreign key to the user table and an indexed ngram username column:
USERSEARCH
user_id username_ngram
-------------------------
1 crazyguy23
1 razyguy23
1 azyguy23
1 zyguy23
...
The query would then be:
SELECT user_id FROM myapp.usersearch WHERE username_ngram LIKE {string}%
LIMIT 10
I am aware that third party solutions exist, but I would like to stay away from them at the moment for other reasons. Is this approach viable in terms of speed? Am I overestimating the power of indexes if the db would need to check all O(30n) rows where n is the number of users?
Probably not. The union distinct is going to process each subquery to completion.
If you just want arbitrary rows, phrase this as:
(SELECT user_id
FROM myapp.usersearch
WHERE username_1 LIKE {string}%
LIMIT 10
) UNION DISTINCT
(SELECT user_id
FROM myapp.usersearch
WHERE username_2 LIKE {string}%
LIMIT 10
)
LIMIT 10;
This will at least save you lots of time for common prefixes -- say 'S'.
That said, this just returns an arbitrary list of 10 user_ids when there might be many more.
I don't know if the speed will be fast enough for your application. You have to make that judgement by testing on an appropriate set of data.
Assuming SSDs, that should be blazing fast, yes.
Here are some further optimizations:
I would add a DISTINCT to your query, since there is no point in returning the same user_id multiple times. Especially when searching for a very common prefix, such as a single letter.
Also consider searching only for at least 3 letters of input. Less tends to be meaningless (since hopefully your usernames are at least 3 characters long) and is a needless hit on your database.
If you're not adding any more columns (I hope you're not, since this table is meant for blazing fast search!), we can do better. Swap the columns. Make the primary key (username_ngram, user_id). This way, you're searching directly on the primary key. (Note the added benefit of the alphabet ordering of the results! Well... alphabetic on the matching suffixes, that is, not the full usernames.)
Make sure you have an index on user_id, to be able to replace everything for a user if you ever need to change a username. (To do so, just delete all rows for that user_id and insert brand new ones.)
Perhaps we can do even better. Since this is just for fast searching, you could use an isolation level of READ_UNCOMMITTED. That avoids placing any read locks, if I'm not mistaken, and should be even faster. It can read uncommitted data, but so what... Afterwards you'll just query any resulting user_ids in another table and perhaps not find them, if that user was still being created. You haven't lost anything. :)
I think you nedd to use mysql full text index to improve performance. You need to change your syntax to use your full text index.
Create full text index:
CREATE FULLTEXT INDEX ix_usersearch_username_ngram ON usersearch(username_ngram);
The official mysql documentation how to use full text index: https://dev.mysql.com/doc/refman/8.0/en/fulltext-search.html