一句T-SQL语句引发的思考转帖

发表于:2007-05-25来源:作者:点击数: 标签:一句T-SQL引发转帖语句
关于MS SQL SERVER索引优化问题: 有表Stress_test(id int, key char(2)) [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]id 上有普通索引; [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]key 上有簇索引; [$nbsp][$nbsp][$nbsp][$nbs

  关于MS SQLSERVER索引优化问题:
  有表Stress_test(id int, key char(2))
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]id 上有普通索引;
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]key 上有簇索引;
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]id 有有限量的重复;
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]key 有无限量的重复;
  
  现在我需要按逻辑与查询表中key='Az' AND key='Bw' AND key='Cv' 的id
  
  求教高手最有效的查询语句
  
  测试环境:
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Hardware:P4 2.6+512M+80G
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Software:windows server 2003(Enterprise Edition)+Sqlserver 2000 +sp3a
  
  [$nbsp][$nbsp]首先我们建立一个测试的数据,为使数据尽量的分布和随即,我们通过RAND()来随机产生2个随机数再组合成一个字符串,首先插入的数据是1,000,000条记录,然后在循环插入到58,000,000条记录。
  [$nbsp][$nbsp][$nbsp]因为是随机产生的数据,所以如果你自己测试的数据集和我测试的会不一样,但对索引的优化和运行的效率是一样的。
  [$nbsp][$nbsp][$nbsp]下面的“--//测试脚本”是产生测试数据的脚本,你可以根据需要修改 @maxgroup, @maxLoop的值,比如测试1百万的记录可以:
  
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Select @maxgroup=1000
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Select @maxLoop=1000
  
  如果要测试5千万:
  
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Select @maxgroup=5000
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]Select @maxLoop=10000
  
  所以如果你的SERVER或PC比较慢,请耐心等待.....,
  (在我的PC上运行的速度是插入1百万条的时间是1.14m,插入5千八百万条的时间是19.41m,重新建立INDEX的时间是34.36m)
  
  
  
  作为一般的开发人员很容易就想到的语句:
  
  [$nbsp][$nbsp][$nbsp]--语句1
  
  [$nbsp][$nbsp][$nbsp][$nbsp]select a.[id] from
  [$nbsp][$nbsp][$nbsp][$nbsp](select distinct [id] from stress_test where [key] = 'Az') a,
  [$nbsp][$nbsp][$nbsp][$nbsp](select distinct [id] from stress_test where [key] = 'Bw') b ,
  [$nbsp][$nbsp][$nbsp][$nbsp](select distinct [id] from stress_test where [key] = 'Cv') c
  [$nbsp][$nbsp][$nbsp][$nbsp]where a.id = b.id and a.id = c.id
  
  [$nbsp][$nbsp][$nbsp]--语句2
  
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select [id]
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]from stress_test
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where [key]='Az' or [key]='Bw' or [key]='Cv'
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]group by id having(count(distinct [key])=3)
  
  [$nbsp][$nbsp][$nbsp]--语句5
  
  [$nbsp][$nbsp][$nbsp][$nbsp]SELECT distinct a.[id] FROM stress_test AS a,stress_test AS b,stress_test AS c
  [$nbsp][$nbsp][$nbsp][$nbsp]WHERE a.[key]='Az' AND b.[key]='Bw' AND c.[key]='Cv'
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]AND a.[id]=b.[id] AND a.[id]=c.[id]
  
  但作为T-SQL的所谓“高手”可能会认为这种写法很“土”,也显得没有水平,所以会选择一些子查询和外连接的写法,按常理子查询的效率是比较高的:
  
  [$nbsp][$nbsp][$nbsp]--语句3
  
  [$nbsp][$nbsp][$nbsp][$nbsp]select distinct [id] from stress_test A where
  [$nbsp][$nbsp][$nbsp][$nbsp]not exists (
  [$nbsp][$nbsp][$nbsp][$nbsp]select 1 from
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp](select 'Az' as k union all select 'Bw' union all select 'Cv') B
  [$nbsp][$nbsp][$nbsp][$nbsp]left join stress_test C on C.id=A.id and B.[k]=C.[key]
  [$nbsp][$nbsp][$nbsp][$nbsp]where C.id is null)
  
  [$nbsp][$nbsp][$nbsp]--语句4
  
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select distinct a.id from stress_test a
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where not exists
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]( select * from keytb c
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where not exists
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]( select * from stress_test b
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]b.id = a.id
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]and
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]c.kf1 = b.[key]
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp])
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp])
  
  我们先分析这几条语句(针对5千8百万条数据进行分析):
  
  请大家要特别留心Estimated row count的值。
  
  语句1:从执行规划中我们可以看出,MSSQLSERVER选择的索引优化非常有规律,先通过CLUSTERED INDEX筛选出符合[KEY]='Az'条件的ID,然后进行HASH MATCH,在找出ID相等的;依次类推最终检索到符合所有条件的记录。中间的Estimated row count的值都不大。
  
  语句2:从执行规划中我们可以看出,是先通过CLUSTERED INDEX筛选出符合 [key]='Az' or [key]='Bw' or [key]='Cv' 符合所有条件的ID,然后分组进行2次HASH MATCH 所有的ID。我们可以看出Estimated row count的值是越来越少,从最初的369,262到最后排序的只有402。
  
  语句3:从执行规划中我们可以看是非常复杂的,是先通过3组 通过CONSTANT SCAN和NON-CLUSTERED INDEX检索出符合 A.ID=C.ID AND [key]='**' 的记录3组,然后分组进行外键匹配,再将3组的数据合并,排序,然后再和一个NON-CLUSTERED INDEX检索出的记录集进行外键匹配,我们可以看出MSSQLSERVER会对所有的记录(5千万条)记录进行分组,Estimated row count的值是:58,720,000,所以这句T-SQL的瓶颈是对5千万条记录进行分组。
  
  语句4:从执行规划中我们可以看和语句3有相似之处,都要对所有的记录(5千万条)记录进行分组,所以这是检索的瓶颈,而且使用的索引都是NON-CLUSTERED INDEX。
  
  语句5:从执行规划中我们可以看出,先通过CLUSTERED INDEX检索出符合[Key]='Az'的记录集,然后进行HASH MATCH和SORTS,因为数量少所以是非常会的,在和通过NON-CLUSTERED INDEX检索[KEY]='Bw'的记录进行INNER JOIN,在和通过CLUSTERED INDEX检索[KEY]='Cv'的记录进行合并,最后是对4百万条数据进行分组检索,如果是6列,我们可以看出Estimated row count的值是递增,越来越大,最后的分组检索的Estimated row count的值是3.46E+15,这已经形成巨大的瓶颈。
  
  我们可以先测试一下小的数据量(50000条);
  
  大家可以下面测试脚本的:
  
  [$nbsp][$nbsp][$nbsp]Select @maxgroup=500
  [$nbsp][$nbsp][$nbsp]Select @maxLoop=100
  
  ----------------------------------------------------------------------
  |------------------语句 1----语句 2----语句 3----语句 4----语句 5----|
  | 5万(3列) 5ms 19ms 37ms 59ms 0ms
  | 5万(6列) 1ms 26ms 36ms 36ms 1ms
  
  
  从测试的的数据来看,语句5的效率是最高的,几乎没有花费时间,而语句2的效率只能说是一般。如果测试到这里就结束了,我们可以毫不犹豫的选择语句 5 :-(,继续进行下面的测试.....
  
  我们测试百万条以上的记录:
  1.先对1百万条记录进行测试(选取3列)
  2.先对1百万条记录进行测试(选取6列)
  3.对5千万条数据测试(选取3列)
  4.对5千万条数据测试(选取6列)
  
  统计表1:
  ----------------------------------------------------------------------
  |------------------语句 1----语句 2----语句 3----语句 4----语句 5----|
  | 1百万(3列) 0.77% 0.41% 49.30% 48.99% 0.52%
  | 1百万(6列) 1.61% 0.81% 48.99% 47.44% 1.14%
  | 5千万(3列) 0.14% 0.18% 48.88% 48.86% 1.93%
  | 5千万(6列) 0.00% 0.00% 0.00% 0.00% 100.00%
  统计表2:
  ----------------------------------------------------------------------
  |------------------语句 1----语句 2----语句 3----语句 4----语句 5----|
  | 1百万(3列) 9ms 22ms 723ms 753ms 4ms
  | 1百万(6列) 15ms 38ms 764ms 773ms 11ms
  | 5千万(3列) 575ms 262ms 110117ms 110601ms 12533ms
  | 5千万(6列) 1070ms 576ms 107988ms 109704ms 10m以上
  
  测试总结:(我们可以比较关注:语句 2和语句 5)
  1.在1百万条记录的情况下,语句 5是最快的,但在5千万条记录下是最慢的。这说明INDEX的优化一定的情况下,数据量不同,检索的效率也是不同的。我们平时在写T-SQL时一般关注的时INDEX的使用,只要我们写的T-SQL是利用CLUSTERED INDEX,我们就认为是最优化了,其实这是一个误区,我们还要关注Estimated row count的值,大量的I/O操作是我们应该关注的,所以我们应该根据数据量的不同选择相应的T-SQL语句,不要认为在小数据量下是最高的在大数据量的状态下也许是最慢的:-(。
  
  2.在执行规划中最快的,并不是运行最快的,我们可以看在1百万(6列)在这行中,语句 2和语句 5的比例是0.81%:1.14%,但实际的运行效率是,38ms:11ms。所以,我们在选择T-SQL是要考虑本地I/O的速度,所以在优化语句时不仅要看执行规划还要计算一下具体的效率。
  
  在测试的语句上加入:
  
  [$nbsp][$nbsp][$nbsp][$nbsp]SET STATISTICS TIME ON/OFF
  [$nbsp][$nbsp][$nbsp][$nbsp]SET STATISTICS IO ON/OFF
  是一个很好的调试方法。
  
  3.综合评价,语句 2的效率是最高的,执行效率没有随数据量变化而有很大的差别。
  
  4.执行规划越简单的语句(语句1),综合效率越高,反之则越低(语句3,语句4)。
  
  5.在平时写T-SQL语句时,一定要根据不同的数据量进行测试,虽然都是用CLUSTERED INDEX,但检索的效率却大相径庭。
  
  --//测试脚本
  USE Northwind
  GO
  if exists(select * from sysobjects where name=N'stress_test' and type='U')
  Drop table stress_test
  GO
  --//定义测试的表stress_test,存放所有的测试数据
  Create table stress_test([id] int,[key] char(2))
  
  GO
  --//插入测试的数据
  Set nocount on
  --//变量定义
  Declare @id int --//Stress_test ID 值
  Declare @key char(2) --//Stress_test [key] 值
  Declare @maxgroup int --//组最大的循环数
  Declare @maxLoop int --//ID最大的循环数
  Declare @tempGroup int --//临时变量
  Declare @tempLoop int --//临时变量
  Declare @tempint1 int --//临时变量
  Declare @tempint2 int --//临时变量
  Declare @rowcount int --//记录事务提交的行数
  
  --//初始化变量
  Select @id=1
  Select @maxgroup=1000
  Select @maxLoop=1000
  Select @tempGroup=1
  Select @tempLoop=1
  Select @key='
  Select @rowcount=0
  
  while @tempLoop<=@maxLoop
  begin
  while @tempGroup<=@maxGroup
  begin
  [$nbsp][$nbsp]select @tempint1=65+convert(int,rand()*50)
  [$nbsp][$nbsp]select @tempint2=65+convert(int,rand()*100)
  [$nbsp][$nbsp]if (@tempint1>=122 or @tempint2>=122)
  [$nbsp][$nbsp][$nbsp][$nbsp]begin
  [$nbsp][$nbsp][$nbsp][$nbsp]select @tempint1=@tempint1-100
  [$nbsp][$nbsp][$nbsp][$nbsp]select @tempint2=@tempint2-100
  [$nbsp][$nbsp][$nbsp]
  [$nbsp][$nbsp][$nbsp][$nbsp]if (@tempint1<=65 or @tempint2<=65)
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]begin
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select @tempint1=@tempint1+57
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select @tempint2=@tempint2+57
  [$nbsp][$nbsp][$nbsp][$nbsp]end
  [$nbsp][$nbsp][$nbsp]end
  [$nbsp][$nbsp]select @key=char(@tempint1)+char(@tempint2)
  [$nbsp][$nbsp]if @rowcount=0
  [$nbsp][$nbsp]begin tran ins
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]insert into stress_test([id],[key])values(@id,@key)
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select @rowcount=@rowcount+1
  [$nbsp][$nbsp]
  [$nbsp][$nbsp][$nbsp]if @rowcount>3000 --//判断当行数达到3000条时,开始提交事务
  [$nbsp][$nbsp][$nbsp]begin
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]commit tran ins
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]select @rowcount=0
  [$nbsp][$nbsp][$nbsp]end
  [$nbsp][$nbsp]
  [$nbsp][$nbsp]select @tempGroup=@tempgroup+1
  end
  if @rowcount>0
  begin
  [$nbsp][$nbsp]commit tran ins
  [$nbsp][$nbsp]select @rowcount=0
  end
  
  select @tempGroup=1
  select @id=@id+1
  select @tempLoop=@tempLoop+1
  end
  GO
  --//删除KEY值为NULL的记录
  delete stress_test where [key]is null
  GO
  --//建立簇索引PK_STRESS
  Create Clustered index pk_stress on stress_test([Key])
  --//建立非簇索引NI_STRESS_ID
  Create NonClustered index NI_stress_id on stress_test([id])
  GO
  --//定义测试的表keytb
  if exists(select * from sysobjects where name=N'keytb' and type='U')
  Drop table keytb
  GO
  create table keytb -----//存放你需要匹配的值的表
  (
  [$nbsp][$nbsp]kf1 varchar(20)
  )
  
  --//存放你需要匹配的值,暂定为三个
  insert into keytb(kf1) values('Az');
  insert into keytb(kf1) values('Bw');
  insert into keytb(kf1) values('Cv');
  
  --insert into keytb(kf1) values('Du');
  --insert into keytb(kf1) values('Ex');
  --insert into keytb(kf1) values('Fy');
  GO
  
  下面我们就开始测试几种T-SQL的INDEX优化问题:
  
  --先对1百万条/1亿条记录进行测试(选取3列)的T-SQL:
  
  PRINT '第一种语句:'
  SET STATISTICS TIME ON
  SET STATISTICS IO ON
  select a.[id] from
  (select distinct [id] from stress_test where [key] = 'Az') a,
  (select distinct [id] from stress_test where [key] = 'Bw') b ,
  (select distinct [id] from stress_test where [key] = 'Cv') c
  where a.id = b.id and a.id = c.id
  GO
  PRINT '第二种语句:'
  select [id]
  from stress_test
  where [key]='Az' or [key]='Bw' or [key]='Cv'
  group by id having(count(distinct [key])=3)
  GO
  PRINT '第三种语句:'
  select distinct [id] from stress_test A where
  not exists (
  select 1 from
  (select 'Az' as k union all select 'Bw' union all select 'Cv') B
  left join stress_test C on C.id=A.id and B.[k]=C.[key]
  where C.id is null)
  GO
  PRINT '第四种语句:'
  select distinct a.id from stress_test a
  where not exists
  ( select * from keytb c
  [$nbsp][$nbsp][$nbsp]where not exists
  [$nbsp][$nbsp][$nbsp]( select * from stress_test b
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]b.id = a.id
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]and
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]c.kf1 = b.[key]
  [$nbsp][$nbsp][$nbsp])
  )
  GO
  PRINT '第五种语句:'
  SELECT distinct a.[id] FROM stress_test AS a,stress_test AS b,stress_test AS c
  WHERE a.[key]='Ac' AND b.[key]='Bb' AND c.[key]='Ca'
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]AND a.[id]=b.[id] AND a.[id]=c.[id]
  
  GO
  SET STATISTICS TIME OFF
  SET STATISTICS IO OFF
  
  --先对1百万条/1亿条记录进行测试(选取6列)的T-SQL:
  PRINT '第一种语句:'
  SET STATISTICS TIME ON
  SET STATISTICS IO ON
  select a.[id] from
  (select distinct [id] from stress_test where [key] = 'Az') a,
  (select distinct [id] from stress_test where [key] = 'Bw') b ,
  (select distinct [id] from stress_test where [key] = 'Cv') c,
  (select distinct [id] from stress_test where [key] = 'Du') d,
  (select distinct [id] from stress_test where [key] = 'Ex') e,
  (select distinct [id] from stress_test where [key] = 'Fy') f
  where a.[id] = b.[id] and a.[id] = c.[id] and a.[id] = d.[id] and a.[id] = e.[id] and a.[id] = f.[id]
  GO
  PRINT '第二种语句:'
  select [id]
  from stress_test
  where [key]='Az' or [key]='Bw' or [key]='Cv' or [Key]='Du'or [Key]='Ex'or [Key]='Fy'
  group by id having(count(distinct [key])=6)
  GO
  PRINT '第三种语句:'
  select distinct [id] from stress_test A where
  not exists (
  select 1 from
  (select 'Az' as k union all select 'Bw' union all select 'Cv'union all select 'Du'union all select 'Ex'union all select 'Fy') B
  left join stress_test C on C.id=A.id and B.[k]=C.[key]
  where C.id is null)
  GO
  PRINT '第四种语句:'
  select distinct a.id from stress_test a
  where not exists
  ( select * from keytb c
  [$nbsp][$nbsp][$nbsp]where not exists
  [$nbsp][$nbsp][$nbsp]( select * from stress_test b
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]where
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]b.id = a.id
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]and
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]c.kf1 = b.[key]
  [$nbsp][$nbsp][$nbsp])
  )
  GO
  PRINT '第五种语句:'
  SELECT distinct a.[id] FROM stress_test AS a,stress_test AS b,stress_test AS c,stress_test AS d,stress_test AS e,stress_test AS f
  WHERE a.[key]='Az' AND b.[key]='Bw' AND c.[key]='Cv' AND d.[key]='Du' AND e.[key]='Ex' AND f.[key]='Fy'
  [$nbsp][$nbsp][$nbsp][$nbsp][$nbsp]and a.[id] = b.[id] and a.[id] = c.[id] and a.[id] = d.[id] and a.[id] = e.[id] and a.[id] = f.[id]
  
  GO
  SET STATISTICS TIME OFF
  SET STATISTICS IO OFF

原文转自:http://www.ltesting.net