Question

一个星期以来，我一直在努力寻找以下Redshift难题的解决方案（认为我对此非常着迷）：

Redshift中有一个事件表（“ event_user_item”），用户可以通过输入出现在event_value列中的项目代码来触发某些项目的事件。

提交失败由event_type序列 PageLoad-ItemCode-ErrorResponse 组成，但此类事件类型不一定是连续的，这意味着每个user_id在它们之间可以有许多其他事件类型。

我要根据3个不同的user_id发布一小段摘录，其中应阐明有关失败提交的相关方案。

ord_num event_type          event_value     user_id     event_datetime
1       PageLoad                            124         03/09/2018 21:48:39
2       ItemCode            LG56731         124         03/09/2018 21:48:53
4       Details1PageLoad                    124         03/09/2018 21:48:56
8       PageLoad                            124         03/09/2018 22:02:23
9       ItemCode            GU07019         124         03/09/2018 22:02:32
10      ErrorResponse       Some message    124         03/09/2018 22:02:32
51      PageLoad                            228         04/09/2018 12:38:30
52      ItemCode            EQ23487         228         04/09/2018 12:38:33
53      ErrorResponse       Some message    228         04/09/2018 12:38:34
54      PageLoad                            304         04/09/2018 15:43:14
55      ItemCode            OB68102         304         04/09/2018 15:43:57
56      ErrorResponse       Some message    304         04/09/2018 15:43:58
57      ItemCode            PB68102         304         04/09/2018 15:44:21
58      ErrorResponse       Some message    304         04/09/2018 15:44:22
59      PageLoad                            304         05/09/2018 11:19:37
60      ItemCode            OB68102         304         05/09/2018 11:20:17
62      Details1PageLoad                    304         05/09/2018 11:20:20

目标：根据每个ItemCode查找每个user_id提交失败的次数。重要的是不要混淆失败提交和成功提交中的项目代码。另外，可能还会有多个相同项目代码的“失败”条目。

我不是Redshift的专家，尤其是其窗口功能，但是我想坚持的第一个想法是LAG函数。为此，我打算识别符合条件的ord_nums序列，例如

ord_num event_type          event_value     user_id event_datetime           error?     sequence
1       PageLoad                            124     03/09/2018 21:48:39     
2       ItemCode            LG56731         124     03/09/2018 21:48:53     
4       Details1PageLoad                    124     03/09/2018 21:48:56     
8       PageLoad                            124     03/09/2018 22:02:23     
9       ItemCode            GU07019         124     03/09/2018 22:02:32     
10      ErrorResponse       Some message    124     03/09/2018 22:02:32     1       8-9-10
51      PageLoad                            228     04/09/2018 12:38:30     
52      ItemCode            EQ23487         228     04/09/2018 12:38:33     
53      ErrorResponse       Some message    228     04/09/2018 12:38:34     1       51-52-53
54      PageLoad                            304     04/09/2018 15:43:14     
55      ItemCode            OB68102         304     04/09/2018 15:43:57     
56      ErrorResponse       Some message    304     04/09/2018 15:43:58     1       54-55-56
57      ItemCode            PB68102         304     04/09/2018 15:44:21     
58      ErrorResponse       Some message    304     04/09/2018 15:44:22     1       54-57-58
59      PageLoad                            304     05/09/2018 11:19:37     
60      ItemCode            OB68102         304     05/09/2018 11:20:17     
62      Details1PageLoad                    304     05/09/2018 11:20:20

因此，按user_id进行计数：

user_id     nr_failed_submissions   
124         1   
228         1   
304         2

但是，从上述数据集和预期结果可以看出，无法预测要倒退多少记录，我需要一个不能放在LAG内的附加条件...

我尝试了很多选择，但都不适合。

非常有用和有见地的帖子

但是直到现在，我还没有设法将它们全部融合到可行的解决方案中。在Redshift中必须有一种方法可以做到这一点？

Answer 1

此查询将创建“时间范围”，其中time1代表该用户的PageLoad事件的时间戳记，time2代表该用户的下一个PageLoad事件的时间戳记：

WITH timeranges AS
(
  SELECT A.user_id,
         A.event_datetime AS time1,
         nvl(MAX(B.event_datetime),'2099-01-01') AS time2
  FROM foo AS A
    LEFT JOIN foo AS B
           ON A.user_id = B.user_id
          AND A.event_datetime < B.event_datetime
          AND A.event_type = B.event_type
  WHERE A.event_type = 'PageLoad'
  GROUP BY A.user_id,
           A.event_datetime
)

此查询基于此，将每个“ ItemCode”事件与其对应的“ PageLoad”的时间戳关联：

SELECT timeranges.time1 AS pageloadtime,
       foo.*
FROM foo
  LEFT JOIN timeranges
         ON foo.event_datetime >= timeranges.time1
        AND foo.event_datetime < timeranges.time2
WHERE foo.event_type = 'ItemCode'

此查询确定是否有'ErrorResponse'事件是否属于这些范围中的每个范围：

SELECT timeranges.time1 AS pageloadtime,
       timeranges.user_id,
       BOOL_OR(foo.event_type = 'ErrorResponse') AS has_error
FROM timeranges
  LEFT JOIN foo
         ON event_datetime > time1
        AND event_datetime < time2
GROUP BY timeranges.time1,
         timeranges.user_id
HAVING has_error;

这应该满足了我们的所有需求-对于每个pageload事件，我们知道（1）该pageload是否有错误，以及（2）我们知道与该有效负载相关的所有ItemCode事件。这两个结果集之间的连接应该可以为我们提供所需的信息。

redshift的特殊性给我尝试直接连接这两个数据集带来了一些麻烦，因此我不得不创建两个临时表。这个格式错误的查询给了我预期的结果：

create temporary table items_per_pageload as 
with timeranges as (select A.user_id, A.event_datetime as time1, nvl(max(B.event_datetime), '2099-01-01') as time2 from event_user_item as A left join event_user_item as B on A.user_id=B.user_id and A.event_datetime < B.event_datetime and A.event_type=B.event_type
where A.event_type='PageLoad' group by A.user_id, A.event_datetime)
select timeranges.time1 as pageloadtime, event_user_item.* from event_user_item left join timeranges on event_user_item.event_datetime>=timeranges.time1 and event_user_item.event_datetime<timeranges.time2 where event_user_item.event_type='ItemCode'

create temporary table pageloads_with_errors as 
with timeranges as (select A.user_id, A.event_datetime as time1, nvl(max(B.event_datetime), '2099-01-01') as time2 from event_user_item as A left join event_user_item as B on A.user_id=B.user_id and A.event_datetime < B.event_datetime and A.event_type=B.event_type
where A.event_type='PageLoad' group by A.user_id, A.event_datetime)
select timeranges.time1 as pageloadtime, timeranges.user_id, bool_or(event_user_item.event_type='ErrorResponse') as has_error from timeranges left join event_user_item on event_datetime > time1 and event_datetime < time2
group by timeranges.time1, timeranges.user_id having has_error;

select count(1), user_id, event_value from (
select items_per_pageload.* from items_per_pageload join pageloads_with_errors on items_per_pageload.user_id = pageloads_with_errors.user_id and items_per_pageload.pageloadtime = pageloads_with_errors.pageloadtime 
) group by user_id, event_value

Answer 2

根据杰森·罗森代尔（Jason Rosendale）的答案1进行的以下方法和查询对我来说是应有的作用，

create temporary table items_per_pageload as 
with timeranges as (
  select A.user_id
    ,A.event_datetime as time1
    ,nvl(max(B.event_datetime), '2099-01-01') as time2
    ,LEAD(A.event_datetime,1) over (partition by A.user_id order by A.event_datetime) as next_load_time 
  from event_user_item as A 
  left join event_user_item as B on A.user_id=B.user_id and A.event_datetime < B.event_datetime and A.event_type=B.event_type
  where A.event_type='PageLoad' 
  group by A.user_id, A.event_datetime
  )
select timeranges.time1 as pageloadtime, event_user_item.* 
from event_user_item left join timeranges on event_user_item.event_datetime>=timeranges.time1 and event_user_item.event_datetime<nvl(timeranges.next_load_time,timeranges.time2) 
where event_user_item.event_type='ItemCode';

create temporary table pageloads_with_errors as 
with timeranges as (
  select A.user_id
    ,A.event_datetime as time1
    ,nvl(max(B.event_datetime), '2099-01-01') as time2
    ,LEAD(A.event_datetime,1) over (partition by A.user_id order by A.event_datetime) as next_load_time 
  from event_user_item as A left join event_user_item as B on A.user_id=B.user_id and A.event_datetime < B.event_datetime and A.event_type=B.event_type
  where A.event_type='PageLoad' 
  group by A.user_id, A.event_datetime
  )
select timeranges.time1 as pageloadtime,timeranges.user_id,bool_or(event_user_item.event_type='ErrorResponse') as has_error 
from timeranges 
left join event_user_item on event_datetime > time1 and event_datetime < nvl(next_load_time,time2)
group by timeranges.time1,timeranges.user_id 
having has_error;

/* final counts */
select count(1), user_id, event_value from (
    select items_per_pageload.* 
    from items_per_pageload 
    join pageloads_with_errors on items_per_pageload.user_id = pageloads_with_errors.user_id and items_per_pageload.pageloadtime = pageloads_with_errors.pageloadtime 
) 
group by user_id, event_value;

Redshift：查找满足条件的前几行以构成一个序列

2 个答案: