Teradata 分区查询...动态跟踪行

Question

我有一个包含以下列和数据的表格。数据描述了某些客户活动时期

cust_id    s_date       e_date
11111    01.03.2014   31.03.2014
11111    10.04.2014   30.04.2014
11111    01.05.2014   10.05.2014
11111    15.06.2014   31.07.2014
22222    01.04.2014   31.05.2014
22222    01.06.2014   30.06.2014
22222    01.07.2014   15.07.2014

我想写一个查询来给出这个结果：

cust_id    s_date       e_date
11111    01.03.2014   10.05.2014
11111    15.06.2014   31.07.2014
22222    01.04.2014   15.07.2014

查询结果的目的是当客户 IN-activity 周期小于 15 天时，将行“合并”为一行。我可以处理“前面的 1 行”，但如果需要合并 3 行或更多行，则它不起作用。我想不出如何编写这个查询。

我查询前的“一半”1 行：

SELECT cust_id
     , start_date     as current_period_start_date
     , end_date       as current_period_end_date
     , end_date+15    as current_period_expired_date
     , coalesce(
            min(current_period_expire_date)
           over(partition by cust_id
                    order by start_date
                     rows between 1 preceding and 1 preceding)
               , cast('1900-01-01' as date)) as previous_period_expire_date
     , case 
         when current_period_start_date <= previous_period_expire_date
         then min(current_period_start_date)
             over(partition by cust_id
                      order by start_date
                       rows between 1 preceding and current row)
         else current_period_start_date
       end as new_current_period_start_date

  FROM MY_DB.my_table
     . . .

另外，是否可以像这样将前置变成动态方式？

... over(partition by ... order by ... rows between X preceding and current row)

Answer 1

我会使用lag() 函数来解决这个问题。此函数可用于识别开始新期间的每一行。然后当这个标志被累积求和时，它提供一个组标识符。下面是代码的样子：

select cust_id, min(s_date) as s_date, max(e_date) as e_date
from (select t.*, sum(GroupStartFlag) over (partition by cust_id order by s_date rows unbounded preceding) as grpid
      from (select cust_id, s_date, e_date,
                   (case when s_date <= lag(e_date) over (partition by cust_id order by s_date) + 15
                         then 0
                         else 1
                    end) as GroupStartFlag
            from  MY_DB.my_table
           ) t
     ) t
group by cust_id, grpid;

注意：Teradata 支持窗口函数，但有时对它们有奇怪的要求。我认为上面的方法可以直接工作，但我没有系统可以测试它。

编辑：

我不确定 Teradata 是否支持 lag() 函数。您可以对相关子查询执行等效操作：

select cust_id, min(s_date) as s_date, max(e_date) as e_date
from (select t.*,
             sum(case when s_date <= prev_edate + 15 then 0 else 1 end) over
                 (partition by cust_id order by s_date rows unbounded preceding) as grpid
      from (select cust_id, s_date, e_date,
                   (select max(e_date) 
                    from MY_DB.my_table t2
                    where t2.cust_id = t.cust_id and
                          t2.s_date < t.s_date
                   ) as prev_edate
            from  MY_DB.my_table t
           ) t
     ) t
group by cust_id, grpid;

Answer 2

Gordon 的答案可以修改，因为基本 LAG 语法很容易重写：

LAG(col, n) OVER (ORDER BY c) 

和一个

一样

MIN(col) OVER (ORDER BY c ROWS BETWEEN n PRECEDING AND n PRECEDING)

可能的默认值作为第三个参数可以使用 COALESCE(LAG...., 默认值) 来完成，只有 IGNORE NULLS 选项非常困难。

这会导致：

SELECT cust_id, MIN(s_date) AS s_date, MAX(e_date) AS e_date
FROM (SELECT t.*, SUM(GroupStartFlag) OVER (PARTITION BY cust_id ORDER BY s_date ROWS UNBOUNDED PRECEDING) AS grpid
      FROM (SELECT cust_id, s_date, e_date,
                   (CASE WHEN s_date <= MIN(e_date) 
                                        OVER (PARTITION BY cust_id 
                                              ORDER BY s_date
                                              ROWS BETWEEN 1 PRECEDING AND 1 PRECEDING) + 15
                         THEN 0
                         ELSE 1
                    END) AS GroupStartFlag
            FROM  vt
           ) t
     ) t
GROUP BY cust_id, grpid;

如果您不需要任何其他列（仅 cust_id 和日期），您还可以使用特定的 TD 13.10 表函数来标准化期间。要包括 15 天的差异，您可以简单地减去/加上 15 天：

WITH cte (cust_id, pd)
AS 
 ( SELECT cust_id, PERIOD(s_date-15, e_date) AS pd
   FROM vt
 )
SELECT cust_id,
   BEGIN(pd)+15,
   END(pd),
   cnt
FROM TABLE (TD_NORMALIZE_OVERLAP_MEET
            (NEW VARIANT_TYPE(cte.cust_id)
                ,cte.pd)
        RETURNS (cust_id INTEGER
                ,pd PERIOD(DATE)
                ,cnt INTEGER) --optional: number of rows normalized in one result row
        HASH BY cust_id
        LOCAL ORDER BY cust_id, pd
        ) AS t;

在 TD 14.10 中，还有一个非常好的语法用于标准化周期：

SELECT cust_id, BEGIN (pd)+15, END(pd) 
FROM
 (
   SELECT NORMALIZE
      cust_id, PERIOD(s_date-15, e_date) AS pd
   FROM vt
 ) AS dt

顺便说一句，周期定义为包含开始但排他结束（即，前一个周期的无间隙周期结束和下一个周期的开始具有相同的值），因此您可能必须将 15 更改为 16 才能获得所需结果。