• full join 横向join ，不能map join 走shuffle
• row_number() over ( partition by 主键 order by $flag desc) rank ... where rank =1 ，走shufle

select 
  id, 
  order_datekey, 
  f_procurement_order, 
from 
  (
    select 
      id, 
      order_datekey, 
      f_procurement_order, 
    
      row_number() over (
        partition by id 
        order by 
          b_flag_i desc
      ) rank 
    from 
      (
        select 
          id, 
          order_datekey, 
          f_procurement_order, 
          0 b_flag_i 
        from 
          ods_pms_procurement_order_item_hm old 
        WHERE 
          c_t >= 1479916800 
          or u_t >= 1479916800 
        union all 
        select 
          id, 
          order_datekey, 
          f_procurement_order, 
          1 b_flag_i 
        from 
          ods_pms_procurement_order_item_hm_delta_64124FEADBFA9720 new
      ) t
  ) st 
where 
  rank = 1;

• 差集 + 并集方式 效率最高 前提是增量数据较少，要不也要走shuffle

SET hive.mapred.mode=nonstrict;
INSERT overwrite TABLE $target.table 
SELECT 
  $stream.format 
FROM 
  $target.table old left anti 
  join ($delta) new on $stream.unique_keys
UNION ALL 
SELECT 
  $stream.format 
FROM 
  $target.table ;

fields = 'id,name'

new = 'new'
old = 'old'
and_str = ' AND '
cmd = []
for field in fields.split(','):
    str = old + '.' + field + ' = ' + new + '.' + field
    cmd.append(str)

print and_str.join(cmd)