SQL server：如何计算每个观察的最大连续变化

Posted 2023-02-16

【中文标题】SQL server：如何计算每个观察的最大连续变化

【英文标题】：SQL server: How to count maximum consercutive change for each observation in

【发布时间】：2018-08-25 09:47:46

【问题描述】：

我使用的是 sql server 2012。我想计算每个 obs 的最大连续更改。像这样的表

snapshot_date	customer_id	Number	Max_consercutive_increase_as_of_each_row
Jan-14	12342	0	0
Feb-14	12342	15	1
Mar-14	12342	45	2
Apr-14	12342	0	2
May-14	12342	15	2
Jun-14	12342	45	2
Jul-14	12342	75	3
Aug-14	12342	105	4
Sep-14	12342	135	5
Oct-14	12342	0	4
Nov-14	12342	0	3
Dec-14	12342	0	2
Jan-15	12342	0	1
Feb-15	12342	0	0
Mar-15	12342	0	0
Apr-15	12342	0	0

从每一行开始，倒数到前面的 06 行（包括当前行）。当然，有些起始行只有 01 或 02 行之前。基于“数字”列增加。在 06 行中，如果最大连续在 2 和 3 之间 --> 我想取 3。

我尝试将游标与 fetch relative -n 行一起使用，但我的代码不起作用。 所以请帮我解决它。

非常感谢！

【问题讨论】：

你能把你的脚本发布到现在你是如何尝试的吗？

我尝试过将游标与 fetch relative 一起使用，例如： fetch relative -5 from Test1cursor into ... fetch relative -5 from Test1cursor into ... end

【参考方案1】：

这应该适合你：

-- Create test data

declare @t table(snapshot_date date,customer_id int,Number int);
insert into @t values ('20140101',12342,    0   ),('20140201',12342,    15  ),('20140301',12342,    45  ),('20140401',12342,    0   ),('20140501',12342,    15  ),('20140601',12342,    45  ),('20140701',12342,    75  ),('20140801',12342,    105 ),('20140901',12342,    135 ),('20141001',12342,    0   ),('20141101',12342,    0   ),('20141201',12342,    0   ),('20150101',12342,    0   ),('20150201',12342,    0   ),('20150301',12342,    0   ),('20150401',12342,    0   );

with d as    -- Add a row number to the dataset
(
    select snapshot_date
            ,customer_id
            ,Number
            ,row_number() over (order by snapshot_date) as rn
    from @t
)
,c as        -- Use a recursive CTE to loop through the dataset and check for increases
(
    select snapshot_date
            ,customer_id
            ,Number
            ,rn
            ,0 as ConsecutiveIncreases
    from d
    where rn = 1

    union all

    select t.snapshot_date
            ,t.customer_id
            ,t.Number
            ,t.rn
            ,case when t.Number > c.Number then c.ConsecutiveIncreases + 1 else 0 end
    from d as t
        join c
            on t.rn = c.rn+1
)
-- Take the MAX consecutive increase where the current row is also an increase,
-- unless the row is not an increase, then subtract the number of non-increases
-- from the MAX consecutive increase to find the number of increases within the last 6 rows.
-- If less than 6 rows to use, just take the MAX increase.
select c.snapshot_date
        ,c.customer_id
        ,c.Number
        ,case when isnull(sum(c2.ConsecutiveIncreases),0) = 0
                then 0
              when count(c2.ConsecutiveIncreases) < 6
                then max(c2.ConsecutiveIncreases)
              else max(c2.ConsecutiveIncreases) - case when c.ConsecutiveIncreases = 0
                                                       then sum(case when c2.ConsecutiveIncreases = 0
                                                                     then 1
                                                                     else 0
                                                                     end
                                                               )
                                                       else 0
                                                       end
          end as MaxConsecutiveIncreases
from c
    left join c as c2
        on c2.rn between c.rn-5 and c.rn
group by c.snapshot_date
        ,c.customer_id
        ,c.Number
        ,c.ConsecutiveIncreases
order by 1

输出：

+---------------+-------------+--------+-------------------------+
| snapshot_date | customer_id | Number | MaxConsecutiveIncreases |
+---------------+-------------+--------+-------------------------+
| 2014-01-01    |       12342 |      0 |                       0 |
| 2014-02-01    |       12342 |     15 |                       1 |
| 2014-03-01    |       12342 |     45 |                       2 |
| 2014-04-01    |       12342 |      0 |                       2 |
| 2014-05-01    |       12342 |     15 |                       2 |
| 2014-06-01    |       12342 |     45 |                       2 |
| 2014-07-01    |       12342 |     75 |                       3 |
| 2014-08-01    |       12342 |    105 |                       4 |
| 2014-09-01    |       12342 |    135 |                       5 |
| 2014-10-01    |       12342 |      0 |                       4 |
| 2014-11-01    |       12342 |      0 |                       3 |
| 2014-12-01    |       12342 |      0 |                       2 |
| 2015-01-01    |       12342 |      0 |                       1 |
| 2015-02-01    |       12342 |      0 |                       0 |
| 2015-03-01    |       12342 |      0 |                       0 |
| 2015-04-01    |       12342 |      0 |                       0 |
+---------------+-------------+--------+-------------------------+

【讨论】：

感谢您的回答。但我认为在 snapshot_date-2014-04-01，Maxconsecutiveincrease 应该是 2，因为前 06 行中的“数字”列增加了 02 倍。请分享更多关于你的逻辑，因为它对我来说是新的

还有一件事@iamdave。我的表大约有 1.000.000 行，所以如果我使用递归 cte 会导致临时数据非常大吗？

@trato 更新了我的脚本。正因为如此，你正在寻找这样做，我还没有看到一种方法以一种有效的基于集合的方式来做到这一点，所以它总是很慢。