Sql server Sql server bug?查询结果为';按表达式分组时不确定?
我有以下疑问Sql server Sql server bug?查询结果为';按表达式分组时不确定?,sql-server,sql-server-2008,Sql Server,Sql Server 2008,我有以下疑问 with cte1 as ( select isnull(A, 'Unknown') as A, isnull(nullif(B, 'NULL'), 'Unknown') as B, C from ... -- uses collate SQL_Latin1_General_CP1_CI_AS when joining group by isnull(A, 'Unknown'), isnull(nullif(
with cte1 as (
select isnull(A, 'Unknown') as A,
isnull(nullif(B, 'NULL'), 'Unknown') as B,
C
from ... -- uses collate SQL_Latin1_General_CP1_CI_AS when joining
group by isnull(A, 'Unknown'), isnull(nullif(B, 'NULL'), 'Unknown'), C
),
cte2 as (select top (2147483647) A, B, C from cte1 order by A, B, C),
-- Removing cte2 makes it work if running directly as SQL query. However,
-- it still behave the same if the code is in view or table function
ctes as (
.... -- pretty complex query joining cte2 multiple times
-- uses row_number(), ntile
)
select count(*) from finalCTE
每次执行时,结果(计数)都会更改。而且比应该的数字要小得多。我发现以下任何一个步骤都可以使它正确
cte1
,并使用具体化表cte1
中将组更改为以下任何形式。
按A分组,isnull(NULL如果(B,'NULL'),'Unknown'),C
按isnull(A,'Unknown')、NULL(B,'NULL')、C分组
按A、NULL(B,'NULL')、C分组
- 在其他CTE中使用
代替cte1
。(更新:此步骤并不总是有效。当它在表函数中时仍然存在问题,尽管直接运行SQL时可以工作)cte2
ALTER function [dbo].[fn] (@para1 char(3))
returns table
return
with cte1 as ( select AAA, BBB, CCC
from dbo.fnBBB(12)
where @para1 = 'xxxx'
union all
select AAA, BBB, CCC
from dbo.fnBBB2(12)
where @para1 = 'yyyy'
),
-- Tested not using cte2, the same behave
cte2 as (select top (2147483647) AAA, BBB, CCC from cte1 order by AAA, BBB, CCC),
t as ( select e.CCC, e.value1, cte2.BBB, cte2.AAA
from dbo.T1 e
join cte2 on e.CCC = cte2.CCC
),
b as ( select BBB, AAA, count(*) count,
case when count(*) / 5 > 10 then 10
else count(*) / 5
end as buckets
from t
group by BBB, AAA
having count(*) >= 5
),
b2
as ( select t.*
from b
cross apply ( select *,
ntile(b.buckets) over ( partition by t.BBB, t.AAA order by value1, CCC )
as bucket
from t
where BBB = b.BBB
and AAA = b.AAA
) t
),
m1
as ( select AAA, BBB, b2.CCC, Date, SId, value2, b2.bucket, --
_asc = row_number() over ( partition by BBB, AAA, bucket, Date, SId order by value2, b2.CCC ),
_desc = row_number() over ( partition by BBB, AAA, bucket, Date, SId order by value2 desc, b2.CCC desc )
,count(*) over (partition by BBB, AAA, bucket, Date, SId) scount
from b2 join dbo.T2 e on b2.CCC = e.CCC
),
median
as ( select BBB, AAA, bucket, Date, SId, avg(value2) value2Median, min(scount) sCount
from m1
where _asc in ( _desc, _desc - 1, _desc + 1 )
group by BBB, AAA, bucket, Date, SId
),
bounds
as ( select BBB, AAA, bucket, min(value1) dboMin, max(value1) value1Max, count(*) count
from b2
group by BBB, AAA, bucket
)
select m.*, b.dboMin, b.value1Max, Count
from median m join bounds b on m.BBB = b.BBB and m.AAA = b.AAA and m.bucket = b.bucket
-- order by BBB, AAA, bucket
cte1中使用的功能:
CREATE function [dbo].[fnBBB](@param int)
returns table
return
with m as ( select * -- only this view has non default collate (..._CS_AS)
from dbo.view1 -- indxed view.
)
select isnull(g.AAA, 'Unknown') as AAA,
isnull(nullif(m1.value, 'NULL'), 'Unknown') as BBB
, m.CCC
from m
left join dbo.mapping m0 on m0.id = 12
and m0.value = m. v1 collate SQL_Latin1_General_CP1_CI_AS
left join dbo.map1 r on r.Country = m0.value
left join dbo.map2 g on g.N = r.N
left join dbo.mapping m1 on m1.id = 20
and m1.value = m.v2 collate SQL_Latin1_General_CP1_CI_AS
where m.run_date > dateadd(mm, -@param, getdate())
group by isnull(g.AAA, 'Unknown'), isnull(nullif(m1.value, 'NULL'), 'Unknown'), m.CCC
SQL是一种基于集合的语言。在这个范例中,返回的行的顺序通常是不相关的。您可以将无序视为默认行为。当您确实希望对行进行排序时,需要在查询中的某处显式使用ORDER BY来指定如何排序 对于正常的无序查询,查询返回的行的实际顺序可能由许多因素决定。例如,磁盘上的行的物理布局、查询优化器实际用于返回行的索引的索引节点顺序、查询计划步骤的实际执行顺序等—其中大部分在执行时确定,甚至可能在后续执行之间有所不同
如果这是您观察到的,那么这根本不是一个bug,而是所有关系数据库引擎中的基本和正常行为。您使用的是一个没有order by的
TOP
子句,对吗?TOP 100%没有任何意义。事实上,优化器只是删除了它。为什么会在那里?让我们做一个实验:删除所有ORDERBY和TOP子句。结果现在稳定了吗?我打赌是的。这不是SQL Server中的错误。你在某处依赖未定义的行为,我们必须找出在哪里。问题仍然存在,所有订单都已删除,所以让我们看看其他地方。你在某处使用排序规则吗?行数或其他排名函数?请发布所有内容,因为这样可以更容易地扫描未定义的行为。如果您认为这是一个bug,请在Microsoft Connect上发布。但几乎可以肯定的是,事实并非如此,你会对这个概念感到困惑。我懒得去看代码墙,你还没有提供一个可运行的测试用例来重现这个问题。@dc7a9163d9-可能99%的时间是关于堆栈溢出的,人们声称他们在SQL Server中发现了一个他们没有发现的错误,并且它工作正常。我对非决定性行为的第一个怀疑是其中一个行数
子句的重复值。例如,BBB、AAA、bucket、Date、SId、value2、b2.CCC的复制。我添加了cte2
以强制表假脱机。但是,即使将其删除,其行为也相同。具体化cte1
解决了问题,通过更改分组也解决了问题。