Hive高阶分析函数

1. GROUPING SETS

示例：实现同一数据集的多重group by操作。事实上GROUPING SETS是多个GROUP BY进行UNION ALL操作的简单表达。

数据：

+----------+----------+---+
|year_month|date      |pv |
+----------+----------+---+
|2015-04   |2015-04-12|10 |
|2015-03   |2015-03-10|2  |
|2015-03   |2015-03-10|9  |
|2015-04   |2015-04-12|4  |
|2015-04   |2015-04-13|5  |
|2015-04   |2015-04-13|6  |
|2015-04   |2015-04-16|7  |
|2015-03   |2015-03-12|3  |
+----------+----------+---+

SQL示例：

SELECT
year_month,
date,
SUM(pv) AS pv,
GROUPING__ID 
FROM data
GROUP BY year_month,date
GROUPING SETS (year_month,date) 
ORDER BY GROUPING__ID;

结果：

+----------+----------+----+------------+
|year_month|date      |pv  |grouping__id|
+----------+----------+----+------------+
|2015-03   |null      |14  |1           |
|2015-04   |null      |32  |1           |
|null      |2015-04-12|14  |2           |
|null      |2015-03-10|11  |2           |
|null      |2015-04-16|7   |2           |
|null      |2015-04-13|11  |2           |
|null      |2015-03-12|3   |2           |
+----------+----------+----+------------+

等价于：

SELECT year_month,NULL,SUM(pv) AS pv,1 AS GROUPING__ID FROM data GROUP BY year_month 
UNION ALL
SELECT NULL,date, SUM(pv) AS pv,2 AS GROUPING__ID FROM data GROUP BY date;

GROUPING__ID，表示结果属于哪一个分组集合

2. CUBE

根据GROUP BY的维度的所有组合进行聚合

SELECT
year_month,
date,
SUM(pv) AS pv,
GROUPING__ID 
FROM data
GROUP BY year_month,date
WITH CUBE
ORDER BY GROUPING__ID;

结果：

+----------+----------+----+------------+
|year_month|date      |pv  |grouping__id|
+----------+----------+----+------------+
|2015-04   |2015-04-12|14  |0           |
|2015-04   |2015-04-16|7   |0           |
|2015-03   |2015-03-10|11  |0           |
|2015-03   |2015-03-12|3   |0           |
|2015-04   |2015-04-13|11  |0           |
|2015-04   |null      |32  |1           |
|2015-03   |null      |14  |1           |
|null      |2015-03-10|11  |2           |
|null      |2015-04-12|14  |2           |
|null      |2015-04-13|11  |2           |
|null      |2015-03-12|3   |2           |
|null      |2015-04-16|7   |2           |
|null      |null      |46  |3           |
+----------+----------+----+------------+

等价于：

SELECT NULL,NULL, SUM(pv) AS pv,0 AS GROUPING__ID FROM data
UNION ALL
SELECT year_month,NULL, SUM(pv) AS pv,1 AS GROUPING__ID FROM data GROUP BY year_month 
UNION ALL
SELECT NULL,date, SUM(pv) AS pv,2 AS GROUPING__ID FROM data GROUP BY date
UNION ALL
SELECT year_month,date, SUM(pv) AS pv,3 AS GROUPING__ID FROM data GROUP BY year_month,date;

3.ROLLUP

CUBE的子集，以最左侧的维度为主，从该维度进行层级聚合。

--比如，以year_month维度进行层级聚合：

SELECT
year_month,
date,
SUM(pv) AS pv,
GROUPING__ID  
FROM data
GROUP BY year_month,date
WITH ROLLUP
ORDER BY GROUPING__ID;

结果：

+----------+----------+----+------------+
|year_month|date      |pv  |grouping__id|
+----------+----------+----+------------+
|2015-03   |2015-03-10|11  |0           |
|2015-03   |2015-03-12|3   |0           |
|2015-04   |2015-04-13|11  |0           |
|2015-04   |2015-04-12|14  |0           |
|2015-04   |2015-04-16|7   |0           |
|2015-04   |null      |32  |1           |
|2015-03   |null      |14  |1           |
|null      |null      |46  |3           |
+----------+----------+----+------------+

可以实现这样的上钻过程：月天的PV -> 月的PV -> 总PV。

--把year_month和date调换顺序，则以date维度进行层级聚合：

SELECT
date,
year_month,
SUM(pv) AS pv,
GROUPING__ID  
FROM data
GROUP BY date,year_month 
WITH ROLLUP
ORDER BY GROUPING__ID

结果：

+----------+----------+----+------------+
|date      |year_month|pv  |grouping__id|
+----------+----------+----+------------+
|2015-04-13|2015-04   |11  |0           |
|2015-04-12|2015-04   |14  |0           |
|2015-04-16|2015-04   |7   |0           |
|2015-03-10|2015-03   |11  |0           |
|2015-03-12|2015-03   |3   |0           |
|2015-04-12|null      |14  |1           |
|2015-03-10|null      |11  |1           |
|2015-04-16|null      |7   |1           |
|2015-04-13|null      |11  |1           |
|2015-03-12|null      |3   |1           |
|null      |null      |46  |3           |
+----------+----------+----+------------+

上面介绍的是Hive中常用的高阶分析函数，下面介绍一些其他实用函数：

4.collect_set 

将分组内的数据放入到一个集合中，具有去重的功能

--统计每个用户具体哪些天访问过

select
  user_id,
  collect_set(visit_date) over(partition by user_id) as visit_date_set 
from wedw_tmp.tmp_url_info;

collect_list：和collect_set一样，但是没有去重功能。

5.sort_array 

数组内排序，通常结合collect_set或者collect_list使用。如collect_list为例子，可以发现日期并不是按照顺序组合的，这里有需求需要按照时间升序的方式来组合。

--按照时间升序来组合

select
  user_id,
  sort_array(collect_list(visit_date) over(partition by user_id)) as visit_date_set 
from wedw_tmp.tmp_url_info;

 6.parse_url 

用于解析url相关的参数，直接上sql。

select
visit_url,
parse_url(visit_url, 'HOST') as url_host, --解析host
parse_url(visit_url, 'PATH') as url_path, --解析path
parse_url(visit_url, 'QUERY') as url_query,--解析请求参数
parse_url(visit_url, 'REF') as url_ref, --解析ref
parse_url(visit_url, 'PROTOCOL') as url_protocol, --解析协议
parse_url(visit_url, 'AUTHORITY') as url_authority,--解析author
parse_url(visit_url, 'FILE') as url_file, --解析filepath
parse_url(visit_url, 'USERINFO') as url_user_info --解析userinfo
from wedw_tmp.tmp_url_info;