How to quickly SELECT 3 random records from a 30k MySQL table with a where filter by a single query?

Question

Well, this is a very old question never gotten real solution. We want 3 random rows from a table with about 30k records. The table is not so big in point of view MySQL, but if it represents products of a store, it's representative. The random selection is useful when one presents 3 random products in a webpage for example. We would like a single SQL string solution that meets these conditions:

1. In PHP, the recordset by PDO or MySQLi must have exactly 3 rows.
2. They have to be obtained by a single MySQL query without Stored Procedure used.
3. The solution must be quick as for example a busy apache2 server, MySQL query is in many situations the bottleneck. So it has to avoid temporary table creation, etc.
4. The 3 records must be not contiguous, ie, they must not to be at the vicinity one to another.

The table has the following fields:

CREATE TABLE Products (
  ID INT(8) NOT NULL AUTO_INCREMENT,
  Name VARCHAR(255) default NULL,
  HasImages INT default 0,
  ...
) ENGINE=InnoDB DEFAULT CHARSET=utf8;

The WHERE constraint is Products.HasImages=1 permitting to fetch only records that have images available to show on the webpage. About one-third of records meet the condition of HasImages=1.

Searching for a Perfection, we first let aside the existent Solutions that have drawbacks:

---

I. This basic solution using ORDER BY RAND(),

is too slow but guarantees 3 really random records at each query:

SELECT ID, Name FROM Products WHERE HasImages=1 ORDER BY RAND() LIMIT 3;

*CPU about 0.10s, scanning 9690 rows because of WHERE clause, Using where; Using temporary; Using filesort, on Debian Squeeze Double-Core Linux box, not so bad but

not so scalable to a bigger table as temporary table and filesort are used, and takes me 8.52s for the first query on the test Windows7::MySQL system. With such a poor performance, to avoid for a webpage isn't-it ?

---

II. The bright solution of riedsio using JOIN ... RAND(),

from MySQL select 10 random rows from 600K rows fast, adapted here is only valid for a single random record, as the following query results in an almost always contiguous records. In effect it gets only a random set of 3 continuous records in IDs:

SELECT Products.ID, Products.Name
FROM Products
INNER JOIN (SELECT (RAND() * (SELECT MAX(ID) FROM Products)) AS ID)
  AS t ON Products.ID >= t.ID
WHERE (Products.HasImages=1)
ORDER BY Products.ID ASC
LIMIT 3;

*CPU about 0.01 - 0.19s, scanning 3200, 9690, 12000 rows or so randomly, but mostly 9690 records, Using where.

---

III. The best solution seems the following with WHERE ... RAND(),

seen on MySQL select 10 random rows from 600K rows fast proposed by bernardo-siu:

SELECT Products.ID, Products.Name FROM Products
WHERE ((Products.Hasimages=1) AND RAND() < 16 * 3/30000) LIMIT 3;

*CPU about 0.01 - 0.03s, scanning 9690 rows, Using where.

Here 3 is the number of wished rows, 30000 is the RecordCount of the table Products, 16 is the experimental coefficient to enlarge the selection in order to warrant the 3 records selection. I don't know on what basis the factor 16 is an acceptable approximation.

We so get at the majority of cases 3 random records and it's very quick, but it's not warranted: sometimes the query returns only 2 rows, sometimes even no record at all.

The three above methods scan all records of the table meeting WHERE clause, here 9690 rows.

A better SQL String?

Answer 1

Ugly, but quick and random. Can become very ugly very fast, especially with tuning described below, so make sure you really want it this way.

(SELECT Products.ID, Products.Name
FROM Products
    INNER JOIN (SELECT RAND()*(SELECT MAX(ID) FROM Products) AS ID) AS t ON Products.ID >= t.ID
WHERE Products.HasImages=1
ORDER BY Products.ID
LIMIT 1)

UNION ALL

(SELECT Products.ID, Products.Name
FROM Products
    INNER JOIN (SELECT RAND()*(SELECT MAX(ID) FROM Products) AS ID) AS t ON Products.ID >= t.ID
WHERE Products.HasImages=1
ORDER BY Products.ID
LIMIT 1)

UNION ALL

(SELECT Products.ID, Products.Name
FROM Products
    INNER JOIN (SELECT RAND()*(SELECT MAX(ID) FROM Products) AS ID) AS t ON Products.ID >= t.ID
WHERE Products.HasImages=1
ORDER BY Products.ID
LIMIT 1)

First row appears more often than it should

If you have big gaps between IDs in your table, rows right after such gaps will have bigger chance to be fetched by this query. In some cases, they will appear significatnly more often than they should. This can not be solved in general, but there's a fix for a common particular case: when there's a gap between 0 and the first existing ID in a table.

Instead of subquery (SELECT RAND()*<max_id> AS ID) use something like (SELECT <min_id> + RAND()*(<max_id> - <min_id>) AS ID)

Remove duplicates

The query, if used as is, may return duplicate rows. It is possible to avoid that by using UNION instead of UNION ALL. This way duplicates will be merged, but the query no longer guarantees to return exactly 3 rows. You can work around that too, by fetching more rows than you need and limiting the outer result like this:

(SELECT ... LIMIT 1)
UNION (SELECT ... LIMIT 1)
UNION (SELECT ... LIMIT 1)
...
UNION (SELECT ... LIMIT 1)
LIMIT 3

There's still no guarantee that 3 rows will be fetched, though. It just makes it more likely.

Answer 2

SELECT Products.ID, Products.Name
FROM Products
INNER JOIN (SELECT (RAND() * (SELECT MAX(ID) FROM Products)) AS ID) AS t ON Products.ID     >= t.ID
WHERE (Products.HasImages=1)
ORDER BY Products.ID ASC
LIMIT 3;

Of course the above is given "near" contiguous records you are feeding it the same ID every time without much regard to the seed of the rand function.

This should give more "randomness"

SELECT Products.ID, Products.Name
FROM Products
INNER JOIN (SELECT (ROUND((RAND() * (max-min))+min)) AS ID) AS t ON Products.ID     >= t.ID
WHERE (Products.HasImages=1)
ORDER BY Products.ID ASC
LIMIT 3;

Where max and min are two values you choose, lets say for example sake:

max = select max(id)
min = 225

Answer 3

This statement executes really fast (19 ms on a 30k records table):

$db = new PDO('mysql:host=localhost;dbname=database;charset=utf8', 'username', 'password');
$stmt = $db->query("SELECT p.ID, p.Name, p.HasImages
                    FROM (SELECT @count := COUNT(*) + 1, @limit := 3 FROM Products WHERE HasImages = 1) vars
                    STRAIGHT_JOIN (SELECT t.*, @limit := @limit - 1 FROM Products t WHERE t.HasImages = 1 AND (@count := @count -1) AND RAND() < @limit / @count) p");
$products = $stmt->fetchAll(PDO::FETCH_ASSOC);

The Idea is to "inject" a new column with randomized values, and then sort by this column. The generation of and sorting by this injected column is way faster than the "ORDER BY RAND()" command.

There "might" be one caveat: You have to include the WHERE query twice.

Answer 4

I've been testing the following bunch of SQLs on a 10M-record, poorly designed database.

SELECT COUNT(ID)
INTO @count
FROM Products
WHERE HasImages = 1;

PREPARE random_records FROM
'(
    SELECT * FROM Products WHERE HasImages = 1 LIMIT ?, 1
) UNION (
    SELECT * FROM Products WHERE HasImages = 1 LIMIT ?, 1
) UNION (
    SELECT * FROM Products WHERE HasImages = 1 LIMIT ?, 1
)';

SET @l1 = ROUND(RAND() * @count);
SET @l2 = ROUND(RAND() * @count);
SET @l3 = ROUND(RAND() * @count);

EXECUTE random_records USING @l1
    , @l2
    , @l3;
DEALLOCATE PREPARE random_records;

It took almost 7 minutes to get the three results. But I'm sure its performance will be much better in your case. Yet if you are looking for a better performance I suggest the following ones as they took less than 30 seconds for me to get the job done (on the same database).

SELECT COUNT(ID)
INTO @count
FROM Products
WHERE HasImages = 1;

PREPARE random_records FROM
'SELECT * FROM Products WHERE HasImages = 1 LIMIT ?, 1';

SET @l1 = ROUND(RAND() * @count);
SET @l2 = ROUND(RAND() * @count);
SET @l3 = ROUND(RAND() * @count);

EXECUTE random_records USING @l1;
EXECUTE random_records USING @l2;
EXECUTE random_records USING @l3;

DEALLOCATE PREPARE random_records;

Bear in mind that both these commands require MySQLi driver in PHP if you want to execute them in one go. And their only difference is that the later one requires calling MySQLi's next_result method to retrieve all three results.

My personal belief is that this is the fastest way to do this.

Answer 5

What about creating another table containing only items with image ? This table will be much lighter as it will contain only one-third of the items the original table has !

------------------------------------------
|ID     | Item ID (on the original table)|
------------------------------------------
|0      | 0                              |
------------------------------------------
|1      | 123                            |
------------------------------------------
            .
            .
            .
------------------------------------------
|10 000 | 30 000                         |
------------------------------------------

You can then generate three random IDs in the PHP part of the code and just fetch'em the from the database.

Answer 6

On the off-chance that you're willing to accept an 'outside the box' type of answer, I'm going to repeat what I said in some of the comments.

The best way to approach your problem is to cache your data in advance (be that in an external JSON or XML file, or in a separate database table, possibly even an in-memory table).

This way you can schedule your performance-hit on the products table to times when you know the server will be quiet, and reduce your worry about creating a performance hit at "random" times when the visitor arrives to your site.

I'm not going to suggest an explicit solution, because there are far too many possibilities on how to build a solution. However, the answer suggested by @ahmed is not silly. If you don't want to create a join in your query, then simply load more of the data that you require into the new table instead.