快速组排名()函数

时间:2017-03-24 02:33:24

标签: mysql performance grouping rank row-number

人们尝试在MySQL中模拟MSSQL RANK()或ROW_NUMBER()函数的方式有很多种,但到目前为止我尝试过的所有函数都很慢。

我有一个看起来像这样的表:

CREATE TABLE ratings
    (`id` int, `category` varchar(1), `rating` int)
;

INSERT INTO ratings
    (`id`, `category`, `rating`)
VALUES
    (3, '*', 54),
    (4, '*', 45),
    (1, '*', 43),
    (2, '*', 24),
    (2, 'A', 68),
    (3, 'A', 43),
    (1, 'A', 12),
    (3, 'B', 22),
    (4, 'B', 22),
    (4, 'C', 44)
;

除了它有220,000条记录。大约有90,000个唯一ID。

我想首先通过查看不是*的类别来对id进行排名,其中较高等级的排名较低。

SELECT g1.id,
       g1.category,
       g1.rating,
       Count(*) AS rank
FROM ratings AS g1
JOIN ratings AS g2 ON (g2.rating, g2.id) >= (g1.rating, g1.id)
AND g1.category = g2.category
WHERE g1.category != '*'
GROUP BY g1.id,
         g1.category,
         g1.rating
ORDER BY g1.category,
         rank

输出:

id  category    rating  rank
2   A   68  1
3   A   43  2
1   A   12  3
4   B   22  1
3   B   22  2
4   C   44  1

然后我想把最小的等级列为id,并将其与*类别中的等级进行平均。提供总查询:

SELECT X1.id,
       (X1.rank + X2.minrank) / 2 AS OverallRank
FROM
  (SELECT g1.id,
          g1.category,
          g1.rating,
          Count(*) AS rank
   FROM ratings AS g1
   JOIN ratings AS g2 ON (g2.rating, g2.id) >= (g1.rating, g1.id)
   AND g1.category = g2.category
   WHERE g1.category = '*'
   GROUP BY g1.id,
            g1.category,
            g1.rating
   ORDER BY g1.category,
            rank) X1
JOIN
  (SELECT id,
          Min(rank) AS MinRank
   FROM
     (SELECT g1.id,
             g1.category,
             g1.rating,
             Count(*) AS rank
      FROM ratings AS g1
      JOIN ratings AS g2 ON (g2.rating, g2.id) >= (g1.rating, g1.id)
      AND g1.category = g2.category
      WHERE g1.category != '*'
      GROUP BY g1.id,
               g1.category,
               g1.rating
      ORDER BY g1.category,
               rank) X
   GROUP BY id) X2 ON X1.id = X2.id
ORDER BY overallrank

给我

id  OverallRank
3   1.5000
4   1.5000
2   2.5000
1   3.0000

这个查询是正确的,我想要的输出,但它只挂在我的220,000条记录的真实表上。我该如何优化它?我的真实表格上有id,ratingcategory以及id,category

的索引

编辑:

SHOW CREATE TABLE ratings的结果:

CREATE TABLE `rating` (
     `id` int(11) NOT NULL,
     `category` varchar(255) NOT NULL,
     `rating` int(11) NOT NULL DEFAULT '1500',
     `rd` int(11) NOT NULL DEFAULT '350',
     `vol` float NOT NULL DEFAULT '0.06',
     `wins` int(11) NOT NULL,
     `losses` int(11) NOT NULL,
     `streak` int(11) NOT NULL DEFAULT '0',
     PRIMARY KEY (`streak`,`rd`,`id`,`category`),
     UNIQUE KEY `id_category` (`id`,`category`),
     KEY `rating` (`rating`,`rd`),
     KEY `streak_idx` (`streak`),
     KEY `category_idx` (`category`),
     KEY `id_rating_idx` (`id`,`rating`)
) ENGINE=InnoDB DEFAULT CHARSET=latin1

PRIMARY KEY是此表查询的最常见用例,这就是它成为群集密钥的原因。值得注意的是,该服务器是具有9GB / s FIO随机读取的固态硬盘10。所以我不怀疑没有聚集的指数会影响很多。

(select count(distinct category) from ratings)的输出为50

为了这可能是数据的方式或对我的疏忽,我被包括在整个表的导出中。它只有200KB压缩:https://www.dropbox.com/s/p3iv23zi0uzbekv/ratings.zip?dl=0

第一个查询需要27秒才能运行

1 个答案:

答案 0 :(得分:0)

您可以使用带有AUTO_INCREMENT列的临时表来生成排名(行号)。

例如 - 为'*'类别生成排名:

drop temporary table if exists tmp_main_cat_rank;
create temporary table tmp_main_cat_rank (
    rank int unsigned auto_increment primary key,
    id int NOT NULL
) engine=memory
    select null as rank, id
    from ratings r
    where r.category = '*'
    order by r.category, r.rating desc, r.id desc;

这可以在30毫秒的时间内运行。使用selfjoin的方法在我的机器上需要45秒。即使(category, rating, id)上有新索引,它仍然需要14秒才能运行。

为每组(每个类别)生成排名有点复杂。我们仍然可以使用AUTO_INCREMENT列,但需要计算并减去每个类别的偏移量:

drop temporary table if exists tmp_pos;
create temporary table tmp_pos (
    pos int unsigned auto_increment primary key,
    category varchar(50) not null,
    id int NOT NULL
) engine=memory
    select null as pos, category, id
    from ratings r
    where r.category <> '*'
    order by r.category, r.rating desc, r.id desc;

drop temporary table if exists tmp_cat_offset;
create temporary table tmp_cat_offset engine=memory
    select category, min(pos) - 1 as `offset`
    from tmp_pos
    group by category;

select t.id, min(t.pos - o.offset) as min_rank
from tmp_pos t
join tmp_cat_offset o using(category)
group by t.id

这在约220毫秒内运行。使用新索引,selfjoin解决方案需要42秒或13秒。

现在您只需要将最后一个查询与第一个临时表结合起来,以获得最终结果:

select t1.id, (t1.min_rank + t2.rank) / 2 as OverallRank
from (
    select t.id, min(t.pos - o.offset) as min_rank
    from tmp_pos t
    join tmp_cat_offset o using(category)
    group by t.id
) t1
join tmp_main_cat_rank t2 using(id);

整体运行时间约为280毫秒而没有额外的索引,而且~240毫秒且索引位于(category, rating, id)

关于selfjoin方法的说明:这是一个优雅的解决方案,并且在小组规模下表现良好。平均组大小<= 2时速度很快。组大小为10时可以接受。但是您的平均组大小为447(count(*) / count(distinct category))。这意味着每行都有447个其他行(平均而言)。您可以通过删除group by子句来查看影响:

SELECT Count(*)
FROM ratings AS g1
JOIN ratings AS g2 ON (g2.rating, g2.id) >= (g1.rating, g1.id)
AND g1.category = g2.category
WHERE g1.category != '*'

结果超过10M行。

然而 - 使用(category, rating, id)上的索引,您的查询在我的计算机上运行33秒。