无论输入参数如何，PERCENTILE_CONT() 都返回相同的值

Question

我想获取表格的第 5、50、95 个百分位数

SELECT col1, col2, col3, AVG(col4), STD(col4), 
    PERCENTILE_CONT(0.05) WITHIN GROUP (ORDER BY col4) 
        OVER (PARTITION BY col1, col2, col3) as 5th_percentile, 
    PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY col4)  
        OVER (PARTITION BY col1, col2, col3) as 50th_percentile, 
    PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY col4)  
        OVER (PARTITION BY col1, col2, col3) as 95th_percentile
FROM table
GROUP BY col1, col2, col3
LIMIT 100

我最终得到的是 5th_percentile == 50th_percentile == 95th_percentile

AVG(col4)   STD(col4)   5th_percentile   50th_percentile  95th_percentile
300.000000  0.000000    300.000000       300.000000       300.000000
67.076600   16.968851   82.031792        82.031792        82.031792
66.166136   11.452172   78.348846        78.348846        78.348846
544.262809  68.269014   605.797302       605.797302       605.797302
22.523138   1.820358    24.000000        24.000000        24.000000

怎么了？

编辑：数据库是 MemSQL

Answer 1

WITH a AS (
SELECT col1, col2, col3, 
        PERCENTILE_CONT(0.05) WITHIN GROUP (ORDER BY col4) 
            OVER (PARTITION BY col1, col2, col3) as 5th_percentile,
        PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY col4) 
            OVER (PARTITION BY col1, col2, col3) as 50th_percentile,
        PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY col4) 
            OVER (PARTITION BY col1, col2, col3) as 95th_percentile
FROM table
)
SELECT DISTINCT col1, col2, col3, 5th_percentile, 50th_percentile, 95th_percentile
FROM a
LIMIT 100

这行得通，看起来你不能用 percentile_cont 进行分组

Answer 2

窗口函数在 GROUP BY 子句之后运行。 GROUP BY 每组生成一行，这就是 PERCENTILE_CONT 窗口函数都返回相同值的原因。

您想先计算窗口函数，然后再计算 GROUP BY。您可以通过将窗口函数放在内部子选择中，将 GROUP BY 放在外部选择中来做到这一点。

这里是来自 postgres 的文档，它解释了窗口函数与 group by 的关系（这是标准的 ANSI SQL，MemSQL 做同样的事情）：

https://www.postgresql.org/docs/current/static/tutorial-window.html

窗口函数考虑的行是查询的 FROM 子句生成的“虚拟表”中​​的行，这些行由其 WHERE、GROUP BY 和 HAVING 子句（如果有）过滤。例如，由于不满足 WHERE 条件而被删除的行不会被任何窗口函数看到。一个查询可以包含多个窗口函数，这些函数通过不同的 OVER 子句以不同的方式对数据进行切片，但它们都作用于由该虚拟表定义的同一行集合。

请注意，在 MemSQL 中，如果您在查询中使用未分组或聚合的列，例如 col4，您会从组中的行中获取任意值，即它的行为类似于 ANY_VALUE 聚合。在 MemSQL 的未来版本中，此查询将改为返回错误，以帮助您避免编写具有此类意外行为的查询。

Answer 3

PERCENTILE_CONT()——至少在某些数据库中——可以是聚合函数或窗口函数。

我认为正在发生的事情是在聚合之后 计算该值——我不知道为什么。老实说，我希望代码会出现语法错误，因为 col4 没有聚合。换句话说，(ORDER BY MAX(col4)) 应该有效，但 (ORDER BY col4) 无效，因为百分位数是在聚合之后计算的。

但请尝试不使用OVER 子句：

SELECT col1, col2, col3, AVG(col4), STD(col4), 
       PERCENTILE_CONT(0.05) WITHIN GROUP (ORDER BY col4)  as 5th_percentile, 
       PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY col4) as 50th_percentile, 
       PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY col4) as 95th_percentile
FROM table
GROUP BY col1, col2, col3
LIMIT 100;

编辑：

您的数据库似乎不支持 PERCENTILE_CONT() 作为聚合函数。没有考虑味道。大多数都这样做。

解决方法是SELECT DISTINCT:

SELECT DISTINCT col1, col2, col3,
       AVG(col4) OVER (PARTITION BY col1, col2, col3),
       STD(col4) OVER (PARTITION BY col1, col2, col3),
       PERCENTILE_CONT(0.05) WITHIN GROUP (ORDER BY col4) OVER (PARTITION BY col1, col2, col3)  as 5th_percentile, 
       PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY col4) OVER (PARTITION BY col1, col2, col3) as 50th_percentile, 
       PERCENTILE_CONT(0.95) WITHIN GROUP (ORDER BY col4) OVER (PARTITION BY col1, col2, col3) as 95th_percentile
FROM table
LIMIT 100;

或者使用子查询。