hive 抽样统计

本文转自http://www.taobaotesting.com/blogs/2468，原文分层抽样的逻辑不是很清楚，按照自己的想法重新实现个


算法中可能会需要做抽样。用hive实现了随机抽样中简单随机、系统和分层抽样的方式，记得抽样的概念还是初中数据接触的
其实很多时候不需要理论，想也是可以想到的，不过还是总结一下
   0.测试表：
  

   drop table songpo_test;
   create table if not exists songpo_test
    (
      refund_id string,
      user_id string,
      cat_id  string,
      cat2_id  string,
      org_id  string,
      gmt_create string
    )
    partitioned by(pt string)
    row format delimited
    fields terminated by ','
    lines terminated by '\n' STORED AS SEQUENCEFILE;

   1. 简单随机抽样 (rand()) 从表中，随机打标，排序，随机抽取100个用户数据
        步骤 1). 给每行记录一个相同的标记
             2). 排序,取前100
       

        sql:
        select *
        from(
            select user_id,flag from (select user_id,'1' as flag from songpo_test) x
            distribute by user_id sort by user_id,flag desc
            )a
        where row_number(user_id)<=100;

   2.系统抽样 mod,rand() 依照userrid取模，分5组，每组随机抽取100个用户，实现如：
                1). 依据user_id，取模，获取 mod_numd
                2). 在mod_num组内然后随机排序，
                3). 从各组取出20条

    

 
       sql:
       select *
        from(
            select refund_id,user_id,mod_num,rank_num from (select refund_id,user_id,cast(10+rand()*100 as double) rank_num,user_id%5 as mod_num from songpo_test) 
            distribute by mod_num sort by mod_num,rank_num desc
            )a
      where row_number(mod_num)<=20;

 
  3.分层抽样   按照每个组的记录数来分层抽样。假设需要抽取EXTRA_NUM条记录
                1). 计算每个分区需要抽记录条数
                2). 在mod_num组内然后随机排序，
                3). 从各组取出cat_extra_num条 
          

 
            drop table test_data_extra_indexs;
            create table test_data_extra_indexs as
            select a.cat_id,cat_num,all_num,cat_num/all_num as extra_lv,(cat_num/all_num)*'EXTRA_NUM'  as cat_extra_num,c.refund_id,c.user_id,c.org_id from 
            (select cat_id,count(1) as cat_num,'1' as key from songpo_test group by cat_id) a 
            join
            (select '1' as key,count(1) as all_num from songpo_test) b 
            on a.key=b.key
            join
            (select * from songpo_test) c 
            on a.cat_id=c.cat_id;


            select *
            from(
              select refund_id,user_id,cat_id,mod_num,rank_num from 
              select refund_id,user_id,cat_id,cast(10+rand()*100 as double) rank_num,user_id%5 as mod_num,cat_extra_num from(
              (select refund_id,user_id,cat_id,cast(10+rand()*100 as double) rank_num,user_id%5 as mod_num from test_data_extra_indexs)  x
              )
              distribute by mod_num sort by mod_num,rank_num desc
            )a
      where row_number(mod_num)<=20;