何谓SQLSERVER参数嗅探 大家听到 嗅探 这个词应该会觉得跟黑客肯定有关系吧,使用工具嗅探一下参数,然后截获,脱裤o(_)o 。 事实上,我觉得大家太敏感了,其实这篇文章跟数据库安全没有什么关系,实际上跟数据库性能调优有关 相信大家有泡SQLSERVER论坛的话
大家听到“嗅探”这个词应该会觉得跟黑客肯定有关系吧,使用工具嗅探一下参数,然后截获,脱裤o(∩_∩)o 。
事实上,我觉得大家太敏感了,其实这篇文章跟数据库安全没有什么关系,实际上跟数据库性能调优有关
相信大家有泡SQLSERVER论坛的话不多不少应该都会见过“参数嗅探”这几个字
这里有三篇帖子都是讲述参数嗅探的
http://social.msdn.microsoft.com/Forums/zh-CN/sqlserverzhchs/thread/caccb7f3-8366-4954-8f8a-145eb6bca9dd
http://msdn.microsoft.com/zh-cn/magazine/ee236412.aspx
http://social.msdn.microsoft.com/Forums/zh-CN/sqlserverzhchs/thread/bfbe54de-ac00-49e9-a83b-f97a60bf74ef
下面我给出一个测试数据库的备份文件,里面有一些表和一些测试数据 ,大家可以去下载,因为我下面用的测试表都是这个数据库里的
只需要还原数据库就可以了,这个数据库是SQL2005版本的,数据库名:AdventureWorks
下面只需要用到三张表,表里面有索引:
[Production].[Product]
[SalesOrderHeader_test]
[SalesOrderDetail_test]
数据库下载链接:AdventureWorks_Full_backup_2013-3-4.bak
其实简单来讲,参数嗅探我的很通俗的解释就是:SQLSERVER用鼻子嗅不到具体参数是多少
所以他不能选择最合适的执行计划去执行你的查询,所以参数嗅探是一个不好的现象。
想真正了解参数嗅探,大家可以先创建下面两个存储过程
存储过程一:
<span> 1</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 2</span> <span>GO</span> <span> 3</span> <span>DROP</span> <span>PROC</span><span> Sniff </span><span> 4</span> <span>GO</span> <span> 5</span> <span>CREATE</span> <span>PROC</span> Sniff(<span>@i</span> <span>INT</span><span>) </span><span> 6</span> <span>AS</span> <span> 7</span> <span>SELECT</span> <span>COUNT</span>(b.<span>[</span><span>SalesOrderID</span><span>]</span>),<span>SUM</span>(p.<span>[</span><span>Weight</span><span>]</span><span>) </span><span> 8</span> <span>FROM</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderHeader_test</span><span>]</span><span> a </span><span> 9</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderDetail_test</span><span>]</span><span> b </span><span>10</span> <span>ON</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span>b.<span>[</span><span>SalesOrderID</span><span>]</span> <span>11</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>Production</span><span>]</span>.<span>[</span><span>Product</span><span>]</span><span> p </span><span>12</span> <span>ON</span> b.<span>[</span><span>ProductID</span><span>]</span><span>=</span>p.<span>[</span><span>ProductID</span><span>]</span> <span>13</span> <span>WHERE</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span><span>@i</span> <span>14</span> <span>GO</span>
存储过程二:
<span> 1</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 2</span> <span>GO</span> <span> 3</span> <span>DROP</span> <span>PROC</span><span> Sniff2 </span><span> 4</span> <span>GO</span> <span> 5</span> <span>CREATE</span> <span>PROC</span> Sniff2(<span>@i</span> <span>INT</span><span>) </span><span> 6</span> <span>AS</span> <span> 7</span> <span>DECLARE</span> <span>@j</span> <span>INT</span> <span> 8</span> <span>SET</span> <span>@j</span><span>=</span><span>@i</span> <span> 9</span> <span>SELECT</span> <span>COUNT</span>(b.<span>[</span><span>SalesOrderID</span><span>]</span>),<span>SUM</span>(p.<span>[</span><span>Weight</span><span>]</span><span>) </span><span>10</span> <span>FROM</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderHeader_test</span><span>]</span><span> a </span><span>11</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderDetail_test</span><span>]</span><span> b </span><span>12</span> <span>ON</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span>b.<span>[</span><span>SalesOrderID</span><span>]</span> <span>13</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>Production</span><span>]</span>.<span>[</span><span>Product</span><span>]</span><span> p </span><span>14</span> <span>ON</span> b.<span>[</span><span>ProductID</span><span>]</span><span>=</span>p.<span>[</span><span>ProductID</span><span>]</span> <span>15</span> <span>WHERE</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span><span>@j</span> <span>16</span> <span>GO</span>
然后请做下面这两个测试
测试一:
<span> 1</span> <span>--</span><span>测试一:</span> <span> 2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 3</span> <span>GO</span> <span> 4</span> <span>DBCC</span><span> freeproccache </span><span> 5</span> <span>GO</span> <span> 6</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>500000</span> <span>--</span><span> int</span> <span> 7</span> <span>--</span><span>发生编译,插入一个使用nested loops联接的执行计划</span> <span> 8</span> <span>GO</span> <span> 9</span> <span>10</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>75124</span> <span>--</span><span> int</span> <span>11</span> <span>--</span><span>发生执行计划重用,重用上面的nested loops的执行计划</span> <span>12</span> <span>GO</span>
测试二:
<span> 1</span> <span>--</span><span>测试二:</span> <span> 2</span> <span> 3</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 4</span> <span>GO</span> <span> 5</span> <span>DBCC</span><span> freeproccache </span><span> 6</span> <span>GO</span> <span> 7</span> <span>SET</span> <span>STATISTICS</span> PROFILE <span>ON</span> <span> 8</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>75124</span> <span>--</span><span> int</span> <span> 9</span> <span>--</span><span>发生编译,插入一个使用hash match联接的执行计划</span> <span>10</span> <span>GO</span> <span>11</span> <span>12</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>50000</span> <span>--</span><span> int</span> <span>13</span> <span>--</span><span>发生执行计划重用,重用上面的hash match的执行计划</span> <span>14</span> <span>GO</span>
从上面两个测试可以清楚地看到执行计划重用的副作用。
由于数据分布差别很大参数50000和75124只对自己生成的执行计划有好的性能,
如果使用对方生成的执行计划,性能就会下降。参数50000返回的结果集比较小,
所以性能下降不太严重。参数75124返回的结果集大,就有了明显的性能下降,两个执行计划的差别有近10倍
对于这种因为重用他人生成的执行计划而导致的水土不服现象,SQSERVERL有一个专有名词,叫“参数嗅探 parameter sniffing”
因为语句的执行计划对变量的值很敏感,而导致重用执行计划会遇到性能问题,就是我上面说的
“
SQLSERVER用鼻子嗅不到具体参数是多少,所以他不能选择最合适的执行计划去执行你的查询
”
本地变量的影响
那对于有parameter sniffing问题的存储过程,如果使用本地变量,会怎样呢?
下面请看测试3。这次用不同的变量值时,都清空执行计划缓存,迫使其重编译
<span>1</span> <span>--</span><span>第一次</span> <span>2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span>3</span> <span>GO</span> <span>4</span> <span>DBCC</span><span> freeproccache </span><span>5</span> <span>GO</span> <span>6</span> <span>SET</span> <span>STATISTICS</span> TIME <span>ON</span> <span>7</span> <span>SET</span> <span>STATISTICS</span> PROFILE <span>ON</span> <span>8</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>50000</span> <span>--</span><span> int</span> <span>9</span> <span>GO</span>
<span>1</span> <span>--</span><span>第二次</span> <span>2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span>3</span> <span>GO</span> <span>4</span> <span>DBCC</span><span> freeproccache </span><span>5</span> <span>GO</span> <span>6</span> <span>SET</span> <span>STATISTICS</span> TIME <span>ON</span> <span>7</span> <span>SET</span> <span>STATISTICS</span> PROFILE <span>ON</span> <span>8</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff</span><span>]</span> <span>@i</span> <span>=</span> <span>75124</span> <span>--</span><span> int</span> <span>9</span> <span>GO</span>
<span>1</span> <span>--</span><span>第三次</span> <span>2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span>3</span> <span>GO</span> <span>4</span> <span>DBCC</span><span> freeproccache </span><span>5</span> <span>GO</span> <span>6</span> <span>SET</span> <span>STATISTICS</span> TIME <span>ON</span> <span>7</span> <span>SET</span> <span>STATISTICS</span> PROFILE <span>ON</span> <span>8</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff2</span><span>]</span> <span>@i</span> <span>=</span> <span>50000</span> <span>--</span><span> int</span> <span>9</span> <span>GO</span>
<span>1</span> <span>--</span><span>第四次</span> <span>2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span>3</span> <span>GO</span> <span>4</span> <span>DBCC</span><span> freeproccache </span><span>5</span> <span>GO</span> <span>6</span> <span>SET</span> <span>STATISTICS</span> TIME <span>ON</span> <span>7</span> <span>SET</span> <span>STATISTICS</span> PROFILE <span>ON</span> <span>8</span> <span>EXEC</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Sniff2</span><span>]</span> <span>@i</span> <span>=</span> <span>75124</span> <span>--</span><span> int</span> <span>9</span> <span>GO</span>
看他们的执行计划:
对于第一句和第二句,因为SQL在编译的时候知道变量的值,所以在做EstimateRows的时候,做得非常准确,选择了最适合他们的执行计划
但是对于第三句和第四句,SQLSERVER不知道@j的值是多少,所以在做EstimateRows的时候,不管代入的@i值是多少,
一律给@j一样的预测结果。所以两个执行计划是完全一样的(都是Hash Match)。
参数嗅探的解决办法
参数嗅探的问题发生的频率并不高,他只会发生在一些表格里的数据分布很不均匀,或者用户带入的参数值很不均匀的情况下。
由于篇幅原因我就不具体说了,只是做一些归纳
(1)用exec()的方式运行动态SQL
如果在存储过程里不是直接运行语句,而是把语句带上变量,生成一个字符串,再让exec()这样的命令做动态语句运行,
那SQL就会在运行到这句话的时候,对动态语句进行编译。
这时SQL已经知道了变量的值,会根据生成优化的执行计划,从而绕过参数嗅探问题
<span> 1</span> <span>--</span><span>例如前面的存储过程Sniff,就可以改成这样</span> <span> 2</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 3</span> <span>GO</span> <span> 4</span> <span>DROP</span> <span>PROC</span><span> NOSniff </span><span> 5</span> <span>GO</span> <span> 6</span> <span>CREATE</span> <span>PROC</span> NOSniff(<span>@i</span> <span>INT</span><span>) </span><span> 7</span> <span>AS</span> <span> 8</span> <span>DECLARE</span> <span>@cmd</span> <span>VARCHAR</span>(<span>1000</span><span>) </span><span> 9</span> <span>SET</span> <span>@cmd</span><span>=</span><span>'</span><span>SELECT COUNT(b.[SalesOrderID]),SUM(p.[Weight]) </span><span>10</span> <span>FROM [dbo].[SalesOrderHeader_test] a </span><span>11</span> <span>INNER JOIN [dbo].[SalesOrderDetail_test] b </span><span>12</span> <span>ON a.[SalesOrderID]=b.[SalesOrderID] </span><span>13</span> <span>INNER JOIN [Production].[Product] p </span><span>14</span> <span>ON b.[ProductID]=p.[ProductID] </span><span>15</span> <span>WHERE a.[SalesOrderID]=</span><span>'</span> <span>16</span> <span>EXEC</span>(<span>@cmd</span><span>+</span><span>@i</span><span>) </span><span>17</span> <span>GO</span>
(2)使用本地变量local variable
(3)在语句里使用query hint,指定执行计划
在select,insert,update,delete语句的最后,可以加一个"option(
对SQLSERVER将要生成的执行计划进行指导。当DBA知道问题所在以后,可以通过加hint的方式,引导
SQL生成一个比较安全的,对所有可能的变量值都不差的执行计划
<span> 1</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 2</span> <span>GO</span> <span> 3</span> <span>DROP</span> <span>PROC</span><span> NoSniff_QueryHint_Recompile </span><span> 4</span> <span>GO</span> <span> 5</span> <span>CREATE</span> <span>PROC</span> NoSniff_QueryHint_Recompile(<span>@i</span> <span>INT</span><span>) </span><span> 6</span> <span>AS</span> <span> 7</span> <span>SELECT</span> <span>COUNT</span>(b.<span>[</span><span>SalesOrderID</span><span>]</span>),<span>SUM</span>(p.<span>[</span><span>Weight</span><span>]</span><span>) </span><span> 8</span> <span>FROM</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderHeader_test</span><span>]</span><span> a </span><span> 9</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>SalesOrderDetail_test</span><span>]</span><span> b </span><span>10</span> <span>ON</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span>b.<span>[</span><span>SalesOrderID</span><span>]</span> <span>11</span> <span>INNER</span> <span>JOIN</span> <span>[</span><span>Production</span><span>]</span>.<span>[</span><span>Product</span><span>]</span><span> p </span><span>12</span> <span>ON</span> b.<span>[</span><span>ProductID</span><span>]</span><span>=</span>p.<span>[</span><span>ProductID</span><span>]</span> <span>13</span> <span>WHERE</span> a.<span>[</span><span>SalesOrderID</span><span>]</span><span>=</span><span>@i</span> <span>14</span> <span>OPTION</span><span>(RECOMPILE) </span><span>15</span> <span>GO</span>
(4)Plan Guide
可以用下面的方法,在原来那个有参数嗅探问题的存储过程“Sniff”上,解决sniffing问题
<span> 1</span> <span>USE</span> <span>[</span><span>AdventureWorks</span><span>]</span> <span> 2</span> <span>GO</span> <span> 3</span> <span>EXEC</span> <span>[</span><span>sys</span><span>]</span>.<span>[</span><span>sp_create_plan_guide</span><span>]</span> <span> 4</span> <span>@name</span><span>=</span>N<span>'</span><span>Guide1</span><span>'</span><span>, </span><span> 5</span> <span>@stmt</span><span>=</span>N<span>'</span><span>SELECT COUNT(b.[SalesOrderID]),SUM(p.[Weight]) </span><span> 6</span> <span>FROM [dbo].[SalesOrderHeader_test] a </span><span> 7</span> <span>INNER JOIN [dbo].[SalesOrderDetail_test] b </span><span> 8</span> <span>ON a.[SalesOrderID]=b.[SalesOrderID] </span><span> 9</span> <span>INNER JOIN [Production].[Product] p </span><span>10</span> <span>ON b.[ProductID]=p.[ProductID] </span><span>11</span> <span>WHERE a.[SalesOrderID]=@i</span><span>'</span><span>, </span><span>12</span> <span>@type</span><span>=</span>N<span>'</span><span>OBJECT</span><span>'</span><span>, </span><span>13</span> <span>@module_or_batch</span><span>=</span>N<span>'</span><span>Sniff</span><span>'</span><span>, </span><span>14</span> <span>@params</span><span>=</span><span>NULL</span><span>, </span><span>15</span> <span>@hints</span><span>=</span>N<span>'</span><span>option(optimize for(@i=75124))</span><span>'</span><span>; </span><span>16</span> <span>GO</span>
对于Plan Guide,他还可以使用在一般的语句调优里
终于搞定了,因为要搞测试数据的原因所以搞了很久啊~~