1、UDF、UDAF、UDTF简介

在Hive中，所有的运算符和用户定义函数，包括用户定义的和内置的，统称为UDF（User-Defined Functions）。如下图所示：

UDF官方文档：https://cwiki.apache.org/confluence/display/Hive/LanguageManual+UDF

其中，用户自定义聚合函数和内置聚合函数统称为UDAF（User-Defined Aggregate Functions），用户自定义表生成函数和内置表生成函数统称为UDTF（User-Defined Table-Generating Functions）

本文将主要通过具体案例详细介绍Hive的内置表生成函数（UDTF）

2、Hive内置UDTF

Hive内置UDTF官方文档：https://cwiki.apache.org/confluence/display/Hive/LanguageManual+UDF#LanguageManualUDF-Built-inTable-GeneratingFunctions%28UDTF%29

2.1、explode(array/map)

功能：列转行

示例：

select explode(array(1,2,3))
select explode(split('1,2,3', ','))'''
col
1
2
3
'''

select explode(map(1,2,3,4))'''
key	value
1	2
3	4
'''

2.2、posexplode(array)

功能：列转行，第一列添加元素索引（从0开始）

示例：

select posexplode(array(1,2,3))'''
pos	val
0	1
1	2
2	3
'''

2.3、stack(n,v1,v2,…,vk)

功能：将k个数据平均转换成n行，即k/n列，k必须是n的整数倍，空值使用NULL

示例：

-- 将9个元素按顺序分成3行3列
with user_log as (select stack (3,'1001', '2021-12-12', 123,'1002', '2021-12-12', 145,'1001', '2021-12-13', 143)as (id, dt, lowcarbon)
)
select * from user_log'''
user_log.id	user_log.dt	user_log.lowcarbon
1001	     2021-12-12	               123
1002	     2021-12-12	               145
1001	     2021-12-13	               143
'''

2.4、lateral view UDTF

功能：UDTF只允许在SELECT后面跟UDTF，不允许在SELECT后跟其他字段，例如：

select 'CN' as country,explode(array(1,2,3))

Hive报错，SparkSQL不报错。lateral view可以解决这个问题

示例1：字符串分割

-- 方式1
with shop as (select '1001' as pid,'1,2,3' as svsunion select '1002' as pid,'4,5,' as svs
)
select pid,svs,sv from shop
lateral view outer explode(split(svs, ',')) tmp_v as sv-- 方式2
select pid,svs,sv from (select * from (select '1001' as pid,'1,2,3' as svsunion select '1002' as pid,'4,5,' as svs) tmp
) shop
lateral view outer explode(split(svs, ',')) tmp_v as sv'''
pid  	svs	   sv
1001	1,2,3	1
1001	1,2,3	2
1001	1,2,3	3
1002	4,5,	4
1002	4,5,	5
1002	4,5,	
'''

方式1和方式2使用lateral view和lateral view outer效果相同，空缺值显示为空字符串''

示例2：数组

-- 方式1
with shop as (select '1001' as pid,array(1,2,3) as svsunion select '1002' as pid,array(4,5,NULL) as svs
)
select pid,svs,sv from shop
lateral view outer explode(svs) tmp_v as sv-- 方式2
select pid,svs,sv from (select * from (select '1001' as pid,array(1,2,3) as svsunion select '1002' as pid,array(4,5,NULL) as svs) tmp
) shop
lateral view outer explode(svs) tmp_v as sv'''
pid	    svs	       sv
1001	[1,2,3]	    1
1001	[1,2,3]	    2
1001	[1,2,3]	    3
1002	[4,5,null]	4
1002	[4,5,null]	5
1002	[4,5,null]	NULL
'''

方式1和方式2使用lateral view和lateral view outer效果相同，空缺值显示为NULL

示例3：数据存在NULL

-- 方式1
with shop as (select '1001' as pid, '1,2,3' as svsunion select '1002' as pid, NULL as svs
)
select pid,svs,sv from shop
lateral view outer explode(split(svs, ',')) tmp_v as sv-- 方式2
select pid,svs,sv from (select * from (select '1001' as pid, '1,2,3' as svsunion select '1002' as pid, NULL as svs) tmp
) shop
lateral view outer explode(split(svs, ',')) tmp_v as sv-- lateral view结果：
'''
pid	    svs	   sv
1001	1,2,3	1
1001	1,2,3	2
1001	1,2,3	3
'''
-- lateral view outer结果：
'''
pid	    svs	   sv
1001	1,2,3	1
1001	1,2,3	2
1001	1,2,3	3
1002	NULL	NULL
'''

方式1和方式2使用lateral view和lateral view outer效果不同，lateral view空缺值数据丢失，lateral view outer空缺值显示为NULL

lateral view [outer]详解见文章：传送门

2.5、json_tuple(json_str,k1,k2,…)

功能：从json字符串中根据key获取对应的value返回

示例：json_tuple()使用见文章：传送门

2.6、parse_url_tuple(url,p1,p2,…)

功能：从url中根据属性property获取对应的value返回

示例：

select parse_url_tuple('http://facebook.com/path1/p.php?k1=v1&k2=v2#Ref1', 'HOST', 'PATH', 'QUERY', 'REF', 'PROTOCOL', 'QUERY:k1', 'QUERY:k2')'''
c0	            c1	            c2	        c3	    c4	    c5	c6
facebook.com	/path1/p.php	k1=v1&k2=v2	Ref1	http	v1	v2
'''

参数详解见：https://help.aliyun.com/zh/maxcompute/user-guide/parse-url-tuple

2.7、inline(array<struct>)

功能：将结构体数组并列分解为多行

示例：

select inline(array(struct('A',18,date '2023-10-01'),struct('B',20,date '2023-11-01'))) as (col1,col2,col3)'''
col1  col2	      col3
A	   18	2023-10-01
B	   20	2023-11-01
'''