1.1 窗户函数的定义
窗口函数可以拆分为【窗口+函数】。窗口函数官网指路:LanguageManual WindowingAndAnalytics - Apache Hive - Apache Software Foundationhttps://cwiki.apache.org/confluence/display/Hive/LanguageManual+WindowingAndAnalytics
- 窗口:定义函数计算范围(窗口函数:针对分组后的数据,从逻辑角度指定计算的范围,并没有从物理上真正的切分,只有group by 是物理分组,真正意义上的分组)
- 函数:定义函数计算逻辑
- sql 执行顺序:
from ->
join ->
on ->
where ->
group by->
with (可以在分组后面加上 with rollup,在分组之后对每个组进行全局汇总) ->
having ->
select 后面的普通字段,聚合函数->
distinct ->
order by ->
limit
- 窗口函数执行顺序:窗口函数是作用于select后的结果集。select 的结果集作为窗口函数的输入,但是位于 distcint 之前。窗口函数的执行结果只是在原有的列中单独添加一列,形成新的列,它不会对已有的行或列做修改。
1.2 窗户函数的语法
** <窗口函数>window_name over ( [partition by 字段...] [order by 字段...] [窗口子句] )**
- window_name:给窗口指定一个别名。
- over:用来指定函数执行的窗口范围,如果后面括号中什么都不写,即over() ,意味着窗口包含满足where 条件的所有行,窗口函数基于所有行进行计算。
- 符号[] 代表:可选项; | : 代表二选一
- partition by 子句: 窗口按照哪些字段进行分组,窗口函数在不同的分组上分别执行。分组间互相独立。
- order by 子句 :每个partition内部按照哪些字段进行排序,如果没有partition ,那就直接按照最大的窗口排序,且默认是按照升序(asc)排列。
- 窗口子句:显示声明范围(不写窗口子句的话,会有默认值)。常用的窗口子句如下:
rows between unbounded preceding and unbounded following; -- 上无边界到下无边界(一般用于求 总和)
rows between unbounded preceding and current row; --上无边界到当前记录(累计值)
rows between 1 preceding and current row; --从上一行到当前行
rows between 1 preceding and 1 following; --从上一行到下一行
rows between current row and 1 following; --从当前行到下一行
** ps: **over()里面有order by子句,但没有窗口子句时 ,即: **<窗口函数> over ( partition by 字段... order by 字段... ),此时窗口子句是有默认值的-> rows between unbounded preceding and current row (上无边界到当前行)**。
此时窗口函数语法:<窗口函数> over ( partition by 字段... order by 字段... ) 等价于
<窗口函数> over ( partition by 字段... order by 字段... rows between unbounded preceding and current row)
需要注意有个特殊情况:**当order by 后面跟的某个字段是有重复行的时候**, <窗口函数> over ( partition by 字段... order by 字段... ) **不写窗口子句的情况下,窗口子句的默认值是:range between unbounded preceding and current row(上无边界到当前相同行的最后一行)**。
因此,遇到order by 后面跟的某个字段出现重复行,且需要计算【上无边界到当前行】,那就需要**手动指定 窗口子句 rows between unbounded preceding and current row **,偷懒省略窗口子句会出问题~
** 总结如下:**
1、窗口子句不能单独出现,必须有order by子句时才能出现。
2、当省略窗口子句时:
a) 如果存在order by则默认的窗口是unbounded preceding and current row --当前组的第一行到当前行,即在当前组中,第一行到当前行
b) 如果没有order by则默认的窗口是unbounded preceding and unbounded following --整个组
**ps:**窗口函数的执行顺序是在where之后,所以如果where子句需要用窗口函数作为条件,需要多一层查询,在子查询外面进行。
【例如】求出登录记录出现间断的用户Id
select
id
from (
select
id,
login_date,
lead(login_date, 1, '9999-12-31')
over (partition by id order by login_date) next_login_date
--窗口函数 lead(向后取n行)
--lead(column1,n,default)over(partition by column2 order by column3) 查询当前行的后边第n行数据,如果没有就为null
from (--用户在同一天可能登录多次,需要去重
select
id,
date_format(`date`, 'yyyy-MM-dd') as login_date
from user_log
group by id, date_format(`date`, 'yyyy-MM-dd')
) tmp1
) tmp2
where datediff(next_login_date, login_date) >=2
group by id;
- 窗口函数本身也有执行顺序:** <窗口函数>over ( partition by order by 窗口子句 )**的执行顺序:over -> partition by -> order by -> 窗口子句 -> 函数
1.3 窗口函数分类
哪些函数可以是窗口函数呢?(放在over关键字前面的)
聚合函数
sum(column) over (partition by .. order by .. 窗口子句);
count(column) over (partition by .. order by .. 窗口子句);
max(column) over (partition by .. order by .. 窗口子句);
min(column) over (partition by .. order by .. 窗口子句);
avg(column) over (partition by .. order by .. 窗口子句);
** 需要注意:**
1.count(*)操作时会统计null值,count(column)会过滤掉null值;
2.事实上除了count(*)计算,剩余的聚合函数例如: max(column),min(column),avg(column),count(column) 函数会过滤掉null值
** ps : 高级聚合函数:**
** collect_list 收集并形成list集合,结果不去重;**
** collect_set 收集并形成set**集合,结果去重;
** 举例:**
--每个月的入职人数以及姓名
select
month(replace(hiredate,'/','-')),
count(*) as cnt,
collect_list(name) as name_list
from employee
group by month(replace(hiredate,'/','-'));
/*
输出结果
month cn name_list
4 2 ["宋青书","周芷若"]
6 1 ["黄蓉"]
7 1 ["郭靖"]
8 2 ["张无忌","杨过"]
9 2 ["赵敏","小龙女"]
*/
排序函数
** rank() 、row_number() 、dense_rank() 函数不支持自定义窗口子句。**
-- 顺序排序——1、2、3
row_number() over(partition by .. order by .. )
-- 并列排序,跳过重复序号——1、1、3(横向加)
rank() over(partition by .. order by .. )
-- 并列排序,不跳过重复序号——1、1、2(纵向加)
dense_rank() over(partition by .. order by .. )
前后函数
-- 取得column列前边的第n行数据,如果存在则返回,如果不存在,返回默认值default
lag(column,n,default) over(partition by order by) as lag_test
-- 取得column列后边的第n行数据,如果存在则返回,如果不存在,返回默认值default
lead(column,n,default) over(partition by order by) as lead_test
头尾函数
---当前窗口column列的第一个数值,如果有null值,则跳过
first_value(column,true) over (partition by ..order by.. 窗口子句)
---当前窗口column列的第一个数值,如果有null值,不跳过
first_value(column,false) over (partition by ..order by.. 窗口子句)
--- 当前窗口column列的最后一个数值,如果有null值,则跳过
last_value(column,true) over (partition by ..order by.. 窗口子句)
--- 当前窗口column列的最后一个数值,如果有null值,不跳过
last_value(column,false) over (partition by ..order by.. 窗口子句)
1.4 前后函数:lag/lead
lead和lag函数,这两个函数一般用于计算差值,上面已介绍其语法。**lag****和****lead****函数不支持自定义窗口子句。**
-- 取得column列前边的第n行数据,如果存在则返回,如果不存在,返回默认值default
lag(column,n,default) over(partition by order by) as lag_test
-- 取得column列后边的第n行数据,如果存在则返回,如果不存在,返回默认值default
lead(column,n,default) over(partition by order by) as lead_test
二、实际案例
2.1 股票的波峰波谷
0 问题描述
求股票的波峰Crest 和 波谷trough
波峰:当天的股票价格大于前一天和后一天
波谷:当天的股票价格小于前一天和后一天
1 数据准备
create table if not exists table2
(
id int comment '股票id',
dt string comment '日期',
price int comment '价格'
)
comment '股票价格波动信息';
insert overwrite table table2 values
(1,'2019-01-01',10001),
(1,'2019-01-03',1001),
(1,'2019-01-02',1001),
(1,'2019-01-04',1000),
(1,'2019-01-05',1002),
(1,'2019-01-06',1003),
(1,'2019-01-07',1004),
(1,'2019-01-08',998),
(1,'2019-01-09',997),
(2,'2019-01-01',1002),
(2,'2019-01-02',1003),
(2,'2019-01-03',1004),
(2,'2019-01-04',998),
(2,'2019-01-05',999),
(2,'2019-01-06',997),
(2,'2019-01-07',996);
2 数据分析
此题容易理解,利用lag()和lead()函数便可以解决。
select
id,
dt,
price,
case
when price > lag_price and price > lead_price then 'crest'
when price < lag_price and price < lead_price then 'trough'
end as price_type
from (
select
id,
dt,
price,
lag(price, 1) over (partition by id order by dt) as lag_price,
lead(price, 1) over (partition by id order by dt) as lead_price
from table2
) tmp1;
3 小结
lead和lag函数一般用于计算当前行与上一行,或者当前行与下一行之间的差值。在用户间断登陆问题中也遇到过此函数。指路:HiveSQL题——用户连续登陆-CSDN博客文章浏览阅读220次,点赞4次,收藏3次。HiveSQL题——用户连续登陆https://blog.csdn.net/SHWAITME/article/details/135900251?spm=1001.2014.3001.5501
2.2 前后列转换(面试题)
0 问题描述
表temp包含A,B 两列,使用SQL对该B列进行处理,形成C列。按照A列顺序,B列值不变,C列累计技术 B列值变化,则C列重新开始计数,如图所示
1 数据准备
with table4 as (
select 2010 as A,1 as B
union all
select 2011 as A,1 as B
union all
select 2012 as A,1 as B
union all
select 2013 as A,0 as B
union all
select 2014 as A,0 as B
union all
select 2015 as A,1 as B
union all
select 2016 as A,1 as B
union all
select 2017 as A,1 as B
union all
select 2018 as A,0 as B
union all
select 2019 as A,0 as B
)
2 数据分析
with table4 as (
select 2010 as A,1 as B
union all
select 2011 as A,1 as B
union all
select 2012 as A,1 as B
union all
select 2013 as A,0 as B
union all
select 2014 as A,0 as B
union all
select 2015 as A,1 as B
union all
select 2016 as A,1 as B
union all
select 2017 as A,1 as B
union all
select 2018 as A,0 as B
union all
select 2019 as A,0 as B
)
select
A,
B,
row_number() over (partition by T order by A) as C
from (
select
A,
B,
--over (order by A) 本质是 :over(order by rows between unbounded preceding and current row )
--省略的是:上无边界到当前行
sum(change) over (order by A) T
from (
select
A,
B,
-- 向上取一行,取不到的记为0
lag(B, 1, 0) over (order by A) as Lag,
case
when B <> lag(B, 1, 0) over (order by A) then 1
else 0
end as change
from table4
) tmp1
) tmp2;
3 小结
lead /lag函数常用于差值计算。
版权归原作者 爱吃辣条byte 所有, 如有侵权,请联系我们删除。