0


HiveSQL题——炸裂函数(explode/posexplode)

一、炸裂函数的知识点

       炸裂函数(一行变多行)本质属于UDTF函数(接收一行数据,输出一行或者多行数据)。

1.1 炸裂函数

  • ** ****explode **

 (1)explode(array<T> a) --> explode针对数组进行炸裂
    语法:lateral view explode(split(a,',')) tmp  as new_column
    返回值:string
    说明:按照分隔符切割字符串,并将数组中内容炸裂成多行字符串
    举例:select student_score from test lateral view explode(split(student_score,',')) tmp as item; 输出结果为:
      student_score        item
      [a,b,c]        =>     a
                            b
                            c
               
 (2)explode(map<k,v> m) --> explode针对map键值对进行炸裂
    举例:select explode(map('a',1,'b',2,'c',3)) as (key,value); 输出结果为:
    得到                 key value
      {a:1,b:2,c:3} =>   a   1
                         b   2
                         c   3
  • posexplode

posexplode和explode之间的区别:posexplode除了返回数据,还会返回该值的下角标。

 (1)posexplode(array<T> a) 
    语法:lateral view posexploed(split(a,',')) tmp as pos,item 
    返回值:string
    说明:按照分隔符切割字符串,并将数组中内容炸裂成多行字符串(炸裂具备下角标 0,1,2,3)
    举例1:select posexplode (array('a','b','c')) as pos,item; 输出结果为:
                  pos  item
      [a,b,c] =>   0     a
                   1     b
                   2     c
    ---------------------------------
    举例2:对student_name进行炸裂,同时也对student_score进行炸裂,且需要保证炸裂后,学生和成绩一一对应,不能错乱。
   lateral view posexplode(split(student_name,',')) tmp1 as student_name_index,student_name
   lateral view posexplode(split(student_score,',')) tmp2 as student_score_index,student_score
  where student_name_index = student_score_index;

1.2 lateral view 侧写视图

官网链接:LanguageManual LateralView - Apache Hive - Apache Software Foundation

  • 定义:lateral view 通常与UDTF配合使用,侧视图的原理是将UDTF的结果构建成一个类似于视图的表,再将原表中的每一行和UDTF函数输出的每一行进行连接,生成一张新的虚拟表。
  • 举例:select id, name, hobbies, hobby from person lateral view explode(hobbies) tmp as hobby; 代码分析: 对原表person中的hobbies列进行炸裂(一行变多行),利用侧视图lateral view对该UDTF产生的记录设置字段名称为hobby, 再将原表中person的一每行与hobby进行连接形成一个虚拟表,命名为tmp。
  • 注意:使用lateral view时侧写视图时,可以对UDTF产生的记录设置字段名称,上述例子为hobby,产生的hobby字段可以用于group by、order by 、limit等语句中,不需要再单独嵌套一层子查询

二、实际案例

2.1 每个学生及其成绩

0 问题描述

根据学生成绩表,计算学生的成绩。

1 数据准备

create table if not exists table10
(
    class    string comment '班级名称',
    student string comment '学生名称',
    score   string comment '学生分数'
)
    comment '学生成绩表';
INSERT overwrite table table10
VALUES ("1班","小A,小B,小C","80,92,70"),
       ("2班","小D,小E","88,62"),
       ("3班","小F,小G,小H","90,97,85");

2 数据分析

思路一:lateral view + explode


select
    class,
    student,
    score,
    student_name,
    student_score
from table10 lateral view explode(split(student, ',')) tmp1 as student_name
         lateral view explode(split(score, ',')) tmp2 as student_score;

**bug:**上面逻辑能跑通,但是学生姓名和学生成绩对应不上,出现错乱,弃用。

正确的代码如下:

思路二:** lateral view + posexplode**


select
    class,
    student,
    score,
    student_name,
    student_score
from table10 lateral view posexplode(split(student, ',')) tmp3 as student_index_st, student_name
         lateral view posexplode(split(score, ',')) tmp4 as student_index_sc, student_score
where student_index_st = student_index_sc;

说明:student_index_st = student_index_sc 的作用:下角标对齐,实现学生和成绩一一对应

3 小结

上述案例的学生成绩表中,【学生姓名】字段和【学生成绩】都是数组类型的字符串,我们需要对两个字段分别炸裂后,实现每个学生与其成绩一一对应,因此需要借助posexlode函数的pos下角标进行约束。(用explode函数无法实现)

2.2 日期交叉问题

0 问题描述

统计每个品牌的总营销天数(营销日期有重叠的地方需要去重

1 数据准备

create table promotion_info
(
    promotion_id string comment '优惠活动id',
    brand        string comment '优惠品牌',
    start_date   string comment '优惠活动开始日期',
    end_date     string comment '优惠活动结束日期'
) comment '各品牌活动周期表';

insert overwrite table promotion_info
values (1, 'oppo', '2021-06-05', '2021-06-09'),
       (2, 'oppo', '2021-06-11', '2021-06-21'),
       (3, 'vivo', '2021-06-05', '2021-06-15'),
       (4, 'vivo', '2021-06-09', '2021-06-21'),
       (5, 'redmi', '2021-06-05', '2021-06-21'),
       (6, 'redmi', '2021-06-09', '2021-06-15'),
       (7, 'redmi', '2021-06-17', '2021-06-26'),
       (8, 'huawei', '2021-06-05', '2021-06-26'),
       (9, 'huawei', '2021-06-09', '2021-06-15'),
       (10, 'huawei', '2021-06-17', '2021-06-21');

2 数据分析

思路一:用带有下标的炸裂函数posexplode将活动区间炸裂成具体的每一天的日期。即:将同一个品牌的所有活动日期都有列出来,再对重叠的日期进行统一去重


select brand,
    count(distinct event_date)
    from
(
    select
    promotion_id,
    brand,
    start_date,
    -- 用 start_date + 下角标pos 
    date_add(start_date,pos) as event_date,
    pos
from (
         select
             promotion_id,
             brand,
             start_date,
             end_date,
             split(space(datediff(end_date, start_date)), '') as ar
         from promotion_info
     ) tmp1
         lateral view posexplode(ar) tmp2 as pos, item
)tmp2
group by brand;

** 思路一的代码拆解分析:**

以一条数据为例,
 promotion_id      brand       start_date       end_date
     1             'oppo'     '2021-06-05'    '2021-06-09'
(1)  split(space(datediff(end_date, start_date)), '') as diff 的结果:
      根据[9-5]=4,利用space函数生成长度是4的空格字符串,再利用split函数切割
       1 (promotion_id) , 'oppo'(brand) , '2021-06-05'(start_date) ,'2021-06-09'(end_date) 
        ,  diff ["","","","",""]

(2)用posexplode经过转换增加行(列转行,炸裂),通过下角标pos来获取 event_date,
     根据数组["","","","",""],得到pos的取值是0,1,2,3,4
     炸裂得出下面五行数据(一行变五行)
     1,oppo,2021-06-05(start_date),2021-06-05= date_add(2021-06-05,0) (event_date= start_date+pos)
     1,oppo,2021-06-05(start_date),2021-06-06= date_add(2021-06-05,1) (event_date= start_date+pos)
     1,oppo,2021-06-05(start_date),2021-06-07 = date_add(2021-06-05,2) (event_date= start_date+pos)
     1,oppo,2021-06-05(start_date),2021-06-07 = date_add(2021-06-05,3) (event_date= start_date+pos)
     1,oppo,2021-06-05(start_date),2021-06-08 = date_add(2021-06-05,4) (event_date= start_date+pos)
     1,oppo,2021-06-05(start_date),2021-06-09 = date_add(2021-06-05,5) (event_date= start_date+pos)

     炸裂的目的:活动的优惠时间段[ '2021-06-05' ,  '2021-06-09' ] 拆分成具体的每一天event_date: '2021-06-05','2021-06-06','2021-06-07','2021-06-08','2021-06-09'
(3)根据品牌brand进行分组,求count(distinct event_date) ,从而得到每品牌的总营销天数(营销日期有重叠的地方已经去重了)

** 思路二:用带有下标的炸裂函数posexplode**


select brand,
    count(distinct event_date)
    from
(
    select
    promotion_id,
    brand,
    start_date,
    date_add(start_date,pos) as event_date,
    pos
from (
         select
             promotion_id,
             brand,
             start_date,
             end_date,
             split(repeat(',',datediff(end_date, start_date)),',') as ar
         from promotion_info
     ) tmp1
         lateral view posexplode(ar) tmp2 as pos, item
)tmp2
group by brand;
 思路二的代码拆解分析:跟思路一的逻辑基本是一样的 ,区别仅在于:用代码       split(repeat(',',datediff(end_date, start_date)),',') as ar 去替换 split(space(datediff(end_date, start_date)), '') as ar

** 思路三的代码逻辑如下:**


select
    brand,
    --对品牌brand分组求sum的原因:同一个用户可能对应多段不交叉的活动
    sum(datediff(end_date, new_start_date) + 1) days 
from (
         select
             brand,
             new_start_date,
             end_date
         from (
                  select
                      brand,
                      --判断逻辑:1.如果max_end_date是null(意味着当前行就是首行,不存在上一行了),直接取start_date
                      --2.如果max_end_date不是null,进一步判断【当前行】的start_date与max_end_date的大小,如果start_date小,那用max_date+ 1的值作为【当前行】的新new_start_date
                      if(max_end_date is null, start_date,
                         if(start_date > max_end_date, start_date, date_add(max_end_date, 1))) new_start_date,
                      end_date
                  from (
                           select
                               brand,
                               start_date,
                               end_date,
                               -- 开窗范围:同一个品牌内部:上无边界到截止到上一行
                               -- 开窗的计算逻辑:max(end_date)  --> 对【上无边界到上一行】的最大结束时间end_date进行标记,再与当前行的起始时间start_date进行比对
                               max(end_date)
                                   over (partition by brand order by start_date rows between unbounded preceding and 1 preceding) max_end_date
                           from promotion_info
                       ) t1
              ) t2
         -- 需要保证每行数据的新的起始时间new_start_date 比 结束时间end_date 小
         where new_start_date < end_date
     ) t3
group by brand;

** ** 思路三:没有用到炸裂函数,关键思想是:当活动的上一个日期区间A 与 当前的日期区间B出现重叠(日期交叉,有重复数据)时,需要将区间B的起始时间改成区间A的结束时间。(修改之后需要保证B区间的结束时间> 开始时间)

3 小结

** 上述代码中用到的函数有:**

一、字符串函数
 1、空格字符串函数:space
 语法:space(int n)
 返回值:string
 说明:返回值是n的空格字符串
 举例:select length (space(10)) --> 10
 一般space函数和split函数结合使用:select split(space(3),'');  -->   ["","","",""]

 
 2、split函数(分割字符串)
 语法:split(string str,string pat)
 返回值:array
 说明:按照pat字符串分割str,会返回分割后的字符串数组
 举例:select split ('abcdf','c') from test; -> ["ab","df"]

 3、repeat:重复字符串
 语法:repeat(string A, int n)
 返回值:string
 说明:将字符串A重复n遍。
 举例:select repeat('123', 3); -> 123123123
 一般repeat函数和split函数结合使用:select split(repeat(',',4),',');  -->  
  ["","","","",""]

二、炸裂函数
 explode 
    语法:lateral view explode(split(a,',')) tmp  as new_column
    返回值:string
    说明:按照分隔符切割字符串,并将数组中内容炸裂成多行字符串
    举例:select student_score from test lateral view explode(split(student_score,',')) 
tmp as student_score
 
posexplode
    语法:lateral view posexploed(split(a,',')) tmp as pos,item 
    返回值:string
    说明:按照分隔符切割字符串,并将数组中内容炸裂成多行字符串(炸裂具备瞎下角标 0,1,2,3)
    举例:select student_name, student_score from test
   lateral view posexplode(split(student_name,',')) tmp1 as student_name_index,student_name
   lateral view posexplode(split(student_score,',')) tmp2 as student_score_index,student_score
   where student_score_index = student_name_index
 
 

2.3 用户消费金额

0 问题描述

变更需求:table11表的第1,4列不表,第2列需要变更为连续日期,第3列需要变更成当日累积消费额

1 数据准备

create table if not exists table11
(
    user_id  string comment '用户标识',
    dt       string comment '消费日期',
    price    string comment '消费金额',
    qs       int comment '用户应存期数'
)
    comment '用户消费详情表';
INSERT overwrite table table11
VALUES ("A","2018-12-21","9439.30",12),
       ("A","2019-03-21","9439.30",12),
       ("A","2019-06-21","9439.30",12),
       ("A","2019-09-21","9439.30",12),
       ("B","2018-12-02","9439.30",10),
       ("B","2019-02-02","9439.30",10),
       ("B","2019-06-02","9439.30",10);

2 数据分析

-- 思路一:利用posexplode函数进行炸裂,同时生成下角标pos,
--将消费区间(一行)炸裂成对应的每天的消费日期(多行)
select
    tmp3.user_id,
    tmp3.event_dt,
   -- sum() over(partition by .. order by .. ) 窗口计算的范围是:上无边界(起始行)到当前行,求消费金额的累积值(order by 后面没有窗口子句的情况下,窗口范围是:上无边界(起始行)到当前行)
    cast(sum(tmp4.price) over (partition by tmp3.user_id order by tmp3.event_dt) as decimal(18, 2)) as price,
    tmp3.max_qs
from (
         select
             user_id,
             add_months(min_dt, pos) as event_dt,
             max_qs,
             pos
         from (
                  select
                      user_id,
                      min(dt ) as min_dt,
                      max(price) max_price,
                      max(qs)    max_qs
                  from table11
                  group by user_id
              ) tmp1 lateral view posexplode(split(space(max_qs), '')) tmp2 as pos, item
     ) tmp3
         left join (select
                        user_id,
                        ds,
                        price
                    from table11) tmp4
                   on tmp3.user_id = tmp4.user_id and tmp3.new_ds = tmp4.ds;

3 小结

利用posexplode的下角标pos进行填补连续。利用sum(price)over(partition by ..order by)进行消费金额的累积值统计(截止到当日)

(1)lateral view posexplode(split(space(max_qs), '')) tmp2 as pos, item;-->对字段 期数ds进行posexplode炸裂,一行变多行,且生成对应的下角标pos

(2)add_months(min_ds, pos) as new_ds; --> 基于min_dt + pos对消费日期 进行填补,组成连续的消费日期区间。

待补充:炸裂的弊端是可能会发生数据膨胀,当数据集小的时候,用炸裂方便,当时数据集大时,需慎用。


本文转载自: https://blog.csdn.net/SHWAITME/article/details/135941286
版权归原作者 吵吵叭火 所有, 如有侵权,请联系我们删除。

“HiveSQL题——炸裂函数(explode/posexplode)”的评论:

还没有评论