hiveSQL开窗函数详解

hive开窗函数

1. 开窗函数概述

窗口函数也称OLAP函数，对数据库进行实时分析处理

1.1 窗口函数分类

• 序号函数：row_number() / rank() / dense_rank()
• 分布函数：percent_rank() / cume_dist()
• 前后函数：lag() / lead()
• 头尾函数：first_val() / last_val()
• 聚合函数+窗口函数：sum() over()、 max()/min() over() 、avg() over()
• 其他函数：nth_value() / nfile()

1.2 窗口函数和普通聚合函数的区别

聚合函数是将多条记录聚合成一条，窗口函数是每条记录都会执行，有几条记录执行完还是几条

窗口函数兼具group by子句的分组功能和order by子句的排序功能，但是partition by 子句不具备group by的汇总功能

2. 窗口函数的基本用法

准备基础数据

CREATE TABLE exam_record (uid int COMMENT '用户ID',exam_id int COMMENT '试卷ID',start_time timestamp COMMENT '开始时间',submit_time timestamp COMMENT '提交时间',score tinyint COMMENT '得分'
) 
COMMENT '考试记录表'
ROW FORMAT DELIMITED
FIELDS TERMINATED BY ','
STORED AS TEXTFILE
TBLPROPERTIES ("skip.header.line.count"="1");INSERT INTO exam_record(uid,exam_id,start_time,submit_time,score) VALUES
(1006, 9003, '2021-09-07 10:01:01', '2021-09-07 10:21:02', 84),
(1006, 9001, '2021-09-01 12:11:01', '2021-09-01 12:31:01', 89),
(1006, 9002, '2021-09-06 10:01:01', '2021-09-06 10:21:01', 81),
(1005, 9002, '2021-09-05 10:01:01', '2021-09-05 10:21:01', 81),
(1005, 9001, '2021-09-05 10:31:01', '2021-09-05 10:51:01', 81),
(1004, 9002, '2021-09-05 10:01:01', '2021-09-05 10:21:01', 71),
(1004, 9001, '2021-09-05 10:31:01', '2021-09-05 10:51:01', 91),
(1004, 9002, '2021-09-05 10:01:01', '2021-09-05 10:21:01', 80),
(1004, 9001, '2021-09-05 10:31:01', '2021-09-05 10:51:01', 80);select * from exam_record;

exam_record.uid	exam_record.exam_id	exam_record.start_time	exam_record.submit_time	exam_record.score
1006	9001	2021-09-01 12:11:01	2021-09-01 12:31:01	89
1006	9002	2021-09-06 10:01:01	2021-09-06 10:21:01	81
1005	9002	2021-09-05 10:01:01	2021-09-05 10:21:01	81
1005	9001	2021-09-05 10:31:01	2021-09-05 10:51:01	81
1004	9002	2021-09-05 10:01:01	2021-09-05 10:21:01	71
1004	9001	2021-09-05 10:31:01	2021-09-05 10:51:01	91
1004	9002	2021-09-05 10:01:01	2021-09-05 10:21:01	80
1004	9001	2021-09-05 10:31:01	2021-09-05 10:51:01	80

2.1 基本用法

窗口函数语法

<窗口函数> over[(partition by <列表清单>)] order by <排序列表清单> [rows between 开始位置 and 结束位置]

窗口函数：指要使用的分析函数，

over(): 用来指定窗口函数的范围，如果括号中什么都不写，则窗口包含where的所有行

select uidscore,sum(score) over() as sum_score
from exam_record;

运行结果

uid	score	sum_score
1006	89	654
1006	81	654
1005	81	654
1005	81	654
1004	71	654
1004	91	654
1004	80	654
1004	80	654

2.2 设置窗口的方法

2.2.1 window_name

给窗口指定一个别名

select uid,score,rank() over my_window_name as rk_num,row_number() over my_window_name as row_num
from exam_record
window my_window_name as (partition by uid order by score);

2.2.2 partition by

select uid,score,sum(score) over(partition by uid) as sum_score
from exam_record;

按照uid进行分组，分别求和

使用row_number()序号函数，表明序号

selectuid,score,row_number() over(partition by uid) as row_num
from exam_record;

2.2.3 order by 子句

按照哪些字段进行排序，窗口函数将按照排序后的记录进行编号

selectuid,score,row_number() over (partition by uid order by score desc) as row_num
from exam_record

单独使用order by uid

selectuid,score,sum(score) over (order by uid desc) as row_num
from exam_record;

单独使用partition by uid

selectuid,score,sum(score) over (partition by uid) as row_num
from exam_record;

partition by进行分组内的求和，分区间独立

order by 对序号相同的进行求和，对序号不同的进行累加求和

单独使用order by score

selectuid,score,sum(score) over (order by score desc) as row_num
from exam_record;

2.2.4 rows指定窗口大小

查看score的平均值

selectuid,score,avg(score) over(order by score desc) as avg_num
from exam_record

按照score降序排列，每一行计算前一行到当前行的score的平均值

selectuid,score,avg(score) over(order by row_score) as avg_num
from(selectuid,score,row_number() over(order by score desc) as row_scorefrom exam_record)res

窗口框架

指定窗口大小，框架是对窗口的进一步分区，框架有两种限定方式：

使用rows语句，通过指定当前行之前或之后的固定数目的行来限制分区中的行数

使用range语句，按照排列序列的当前值，根据相同值来确定分区中的行数

order by 字段名 range|rows 边界规则0 | [between 边界规则1] and 边界规则2 

range和rows的区别

range按照值的范围进行范围的定义，rows按照行的范围进行范围的定义

• 使用框架时，必须要有order by子句，如果仅指定了order by子句未指定框架，则默认框架会使用range unbounded preceding and current row （从第一行到当前行的数据）
• 如果窗口函数没有指定order by子句，就不存在 rows|range 窗口的计算
• range 只支持使用unbounded 和 current row

查询我与前两名的平均值

selectuid,score,avg(score) over(order by score desc rows 2 preceding) as avg_score
from exam_record;

查询当前行及前后一行的平均值

selectuid,score,avg(score) over(order by score desc rows between 1 preceding and 1 following) as avg_score
from exam_record;

2.3 开窗函数中加 order by 和不加 order by 的区别

当开窗函数为排序函数时，如row_number()、rank()等，over中的order by 只起到窗口内排序的作用

当开窗函数为聚合函数时，如max、min、count等，over中的order by不仅对窗口内排序，还起到窗口内从当前行到之前所有行的聚合

selectuid,exam_id,start_time,sum(score) over(partition by uid) as one,sum(score) over(partition by uid order by start_time) as two
from exam_record

3. 窗口函数用法举例

3.1 序号函数： row_number() / rank() / dese_rank()

区别：rank() : 并列排序，跳过重复序号------1、1、3

​ row_number() : 顺序排序——1、2、3

​ dese_rank() : 并列排序，不跳过重复序号——1、1、2

selectuid,score,rank() over my_window as rk_num,row_number() over my_window as row_num
from exam_record
window my_window as (partition by uid order by score);

不使用窗口函数实现分数排序

SELECTP1.uid,P1.score,(SELECTCOUNT(P2.score)FROM exam_record P2WHERE P2.score > P1.score) + 1 AS rank_1
FROM exam_record P1
ORDER BY rank_1;

3.2 分布函数: percent_rank() / cume_dist()

3.2.1 percent_rank()

percent_rank() 函数将某个数据在数据集的排位作为数据集的百分比值返回，范围0到1，

按照(rank - 1) / (rows - 1)进行计算，rank为rank()函数产生的序号，rows为当前窗口的记录总行数

selectuid,score,rank() over my_window as rank_num,percent_rank() over my_window as prk
from exam_record
window my_window as (order by score desc)

3.2.2 cume_dist()

如果升序排列，则统计：小于等于当前值的行数 / 总行数

如果降序排列，则统计：大于等于当前值的行数 / 总行数

查询小于等于当前score的比例

selectuid,score,rank() over my_window as rank_num,cume_dist() over my_window as cume
from exam_record
window my_window as (order by score asc);

3.2.3 前后函数lag(expr, n, defval) 、 lead(expr, n, defval)

lag()和lead()函数可以在同一次查询中取出同一字段前 n 行的数据和后 n 行的数据作为独立列

lag( exp_str,offset,defval) over(partition by .. order by …)lead(exp_str,offset,defval) over(partition by .. order by …)

• exp_str 是字段名
• offset是偏移量，即 n 的值
• defval默认值，如何当前行向前或向后 n 的位置超出表的范围，则会将defval的值作为返回值，默认为NULL

查询前1名同学和后一名同学的成绩和当前同学成绩的差值

• 先将前一名、后一名以及当前行的分数放在一起

selectuid,score,lag(score, 1, 0) over my_window as `before`,lead(score, 1, 0) over my_window as `next`
from exam_record
window my_window as (order by score desc);

• 然后做差值

selectuid,score,score - before as before,score - next as next
from (selectuid,score,lag(score, 1, 0) over my_window as before,lead(score, 1, 0) over my_window as next
from exam_record
window my_window as (order by score desc))res

3.2.4 头尾函数：first_value(expr) 、 last_value(expr)

• 返回第一个expr：first_value(expr)
• 返回第二个expr：last_value(expr)

查询第一个和最后一个分数

selectuid,score,first_value(score) over my_window as first,last_value(score) over my_window as last
from exam_record
window my_window as (order by score desc);

4 聚合函数+窗口函数

窗口函数在where之后执行，所以where需要用窗口函数作为条件

 SELECTuid,score,sum(score) OVER my_window_name AS sum_score,max(score) OVER my_window_name AS max_score,min(score) OVER my_window_name AS min_score,avg(score) OVER my_window_name AS avg_scoreFROM exam_recordWINDOW my_window_name AS (ORDER BY score desc)