HiveSQL题——数据炸裂和数据合并

一、数据炸裂

0 问题描述

如何将字符串1-5,16,11-13,9" 扩展成 "1,2,3,4,5,16,11,12,13,9" 且顺序不变。

1 数据准备

with data as (select '1-5,16,11-13,9' as a)

2 数据分析

** 步骤一**：explode(split(a, ',')) 炸裂 + row_number()排序，一行变多行，且对每行的数据排序，保证有序性。

with data as (select '1-5,16,11-13,9' as a)
select
    a1,
    row_number() over () as rn
from (
         select
             explode(split(a, ',')) as a1
         from data
     ) tmp1;

输出结果：

步骤二： lateral view explode(split(a, '-')) 、max(b) - min(b) as diff

（1）lateral view +explode 侧写和炸裂，一行变多行，并将源表中每行的输出结果与该行连接；

（2）group by a1, rn ....... select min(a2) as start_data得到每个分组的起始值

（3）max(a2) - min(a1) 得到每个分组的步长

with data as (select '1-5,16,11-13,9' as a)
select
    a1,
    rn,
    cast(min(a2) as int)           as start_data,
    cast(max(a2) - min(a2) as int) as diff
from (
         select
             a1,
             a2,
             rn
         from (
                  select
                      a1,
                      row_number() over () as rn
                  from (
                           select
                               explode(split(a, ',')) as a1
                           from data
                       ) tmp1
              ) tmp2
                  lateral view explode(split(a1, '-')) table1 as a2
     ) tmp3
group by a1, rn;

输出结果是：

步骤三： 根据步长生成索引值，起始值加上索引值获取展开值

侧写和炸裂，根据分组的步长 diff  生成对应的索引值pos
 （1）lateral view posexplode(split(repeat(',', diff), ',')) table2 as pos, item；
该代码等价于：  lateral view posexplode(split(space(diff), '')) table2 as pos, item;

 （2）(start_data + pos) as  end_data,起始值加上索引值获取展开值

with data as (select '1-5,16,11-13,9' as a)
select
    a1,
    rn,
    start_data,
    diff,
    (start_data + pos) as end_data
from (
         select
             a1,
             rn,
             cast(min(a2) as int)           as start_data,
             cast(max(a2) - min(a2) as int) as diff
         from (
                  select
                      a1,
                      a2,
                      rn
                  from (
                           select
                               a1,
                               row_number() over () as rn
                           from (
                                    select
                                        explode(split(a, ',')) as a1
                                    from data
                                ) tmp1
                       ) tmp2
                           lateral view explode(split(a1, '-')) table1 as a2
              ) tmp3
         group by a1, rn
     ) tmp4
         lateral view posexplode(split(repeat(',', diff), ',')) table2 as pos, item

输出结果是：

步骤四(关键的一步) : row_number() over (order by rn,end_data) as rn1 需要进行全局排序，不然最后collect_list聚合的时候，数字顺序是混乱的。

with data as (select '1-5,16,11-13,9' as a)
select
    a1,
    rn,
    end_data,
    row_number() over (order by rn,end_data) as rn1
from (
         select
             a1,
             rn,
             start_data,
             diff,
             (start_data + pos) as end_data
         from (
                  select
                      a1,
                      rn,
                      cast(min(a2) as int)           as start_data,
                      cast(max(a2) - min(a2) as int) as diff
                  from (
                           select
                               a1,
                               a2,
                               rn
                           from (
                                    select
                                        a1,
                                        row_number() over () as rn
                                    from (
                                             select
                                                 explode(split(a, ',')) as a1
                                             from data
                                         ) tmp1
                                ) tmp2
                                    lateral view explode(split(a1, '-')) table1 as a2
                       ) tmp3
                  group by a1, rn
              ) tmp4
                  lateral view posexplode(split(repeat(',', diff), ',')) table2 as pos, item
     ) tmp5
order by rn

步骤五 : concat_ws(',', collect_list(cast(end_data as string)))

   对end_data 进行聚合collect_list，再利用concat_ws函数拼接字符串

with data as (select '1-5,16,11-13,9' as a)
select
    concat_ws(',', collect_list(cast(end_data as string))) as result
from (
         select
             a1,
             rn,
             end_data,
             row_number() over (order by rn,end_data) as rn1
         from (
                  select
                      a1,
                      rn,
                      start_data,
                      diff,
                      (start_data + pos) as end_data
                  from (
                           select
                               a1,
                               rn,
                               cast(min(a2) as int)           as start_data,
                               cast(max(a2) - min(a2) as int) as diff
                           from (
                                    select
                                        a1,
                                        a2,
                                        rn
                                    from (
                                             select
                                                 a1,
                                                 row_number() over () as rn
                                             from (
                                                      select
                                                          explode(split(a, ',')) as a1
                                                      from data
                                                  ) tmp1
                                         ) tmp2
                                             lateral view explode(split(a1, '-')) table1 as a2
                                ) tmp3
                           group by a1, rn
                       ) tmp4
                           lateral view posexplode(split(repeat(',', diff), ',')) table2 as pos, item
              ) tmp5
         order by rn
     ) tmp6

3 小结

数据炸裂的思路一般是：
    1.计算区间【a,b】的步长(差值)diff；
    2.利用split分割函数+ posexplode等 将一行变成 diff+1 行，生成对应的下角标pos(pos的取值为【0,diff】)；
    3.【a,b】区间的起始值 （a + pos） 将数据平铺开；
    4.基于平铺开后的数据集进一步加工处理，例如：分组聚合等。

上述题目需要注意：collect_list()的底层是ArrayList 集合来实现的，HiveSQL执行时，底层会转换成MR任务，当同时开启多个mapper任务时，无法知道是哪个mapper先计算完，所以可能会出现ArrayList集合中的数据顺序与原来数据插入的顺序不对齐的情况。

关于collect_set() /collect_list() 函数的有序性，相关文章指路：

HiveSQL题——collect_set()/collect_list()聚合函数-CSDN博客HiveSQL题——collect_set()/collect_list()聚合函数https://blog.csdn.net/SHWAITME/article/details/136011647?spm=1001.2014.3001.5502

二、数据合并

0 问题描述

面试题：基于A表的数据生成B表数据

1 数据准备

create table if not exists  tableA
(
    id        string comment '用户id',
    name   string comment '用户姓名'
) comment 'A表';

insert overwrite table tableA values
    ('1','aa'),
    ('2','aa'),
    ('3','aa'),
    ('4','d'),
    ('5','c'),
    ('6','aa'),
    ('7','aa'),
    ('8','e'),
    ('9','f'),
    ('10','g');

create table if not exists  tableC
(
    id     string comment '用户id',
    name   string comment '用户姓名'
) comment 'C表';

insert overwrite table tableC values
    ('3','aa|aa|aa'),
    ('4','d'),
    ('5','c'),
    ('7','aa|aa'),
    ('8','e'),
    ('9','f'),
    ('10','g');

2 数据分析

步骤1：寻找满足条件的断点

select
    id,
    name,
    if(name != lag_name, 1, 0) as flag
from (
         select
             id,
             name,
             lag(name, 1, name) over (order by cast(id as int)) as lag_name
         from tableA
     ) tmp1;

输出结果为:

步骤2：断点处标记为1，非断点处标记为0，并对断点标记值进行累加，构造分组标签

select
    id,
    name,
    --并对断点标记值进行累加，构造分组标签
    sum(flag) over (order by cast(id as int)) grp
from (
         select
             id,
             name,
             --断点处标记为1，非断点处标记为0
             if(name != lag_name, 1, 0) flag
         from (
                  select
                      id,
                      name,
                      lag(name, 1, name) over (order by cast(id as int)) as lag_name
                  from tableA
              ) tmp1
     ) tmp2;

输出结果为：

步骤3：按照分组标签进行数据合并，并取得分组中最大值作为id

select
    max_id,
-- collect_list 数据聚合并拼接concat_ws
    concat_ws('|', collect_list(name)) as name
from (
         select
             name,
             grp,
             max(id) over (partition by grp) max_id
         from (
                  select
                      id,
                      name,
                      sum(if(name != lag_name, 1, 0)) over (order by cast(id as int)) as grp
                  from (
                           select
                               id,
                               name,
                               lag(name, 1, name) over (order by cast(id as int)) as lag_name
                           from tableA
                       ) tmp1
              ) tmp2
     ) tmp3
group by max_id, grp;

输出结果为：

通过max_id, grp分组，对name进行 concat_ws('|', collect_list(name)) 聚合拼接，得出最终的结果

3 小结

 断点分组问题的算法总结
 步骤1：寻找满足条件的断点
 步骤2：断点处标记值为1，非断点处标记为0
 步骤3：对断点标记值进行累加 sum(xx)over(order by xx)，构造分组标签
 步骤4：按照分组标签进行分组求解问题