sqlserver 删除大数据

一、写在前面 - 想说爱你不容易 为了升级数据库至SQL Server 2008 R2，拿了一台现有的PC做测试，数据库从正式库Restore(3个数据库大小夸张地达到100G)，而机器内存只有可怜的4G，不仅要承担DB Server角色，同时也要作为Web Server，可想而知这台机器的命运是

一、写在前面 - 想说爱你不容易

　　为了升级数据库至SQL Server 2008 R2，拿了一台现有的PC做测试，数据库从正式库Restore(3个数据库大小夸张地达到100G+)，而机器内存只有可怜的4G，不仅要承担DB Server角色，同时也要作为Web Server，可想而知这台机器的命运是及其惨烈的，只要MS SQL Server一启动，内存使用率立马飙升至99%。没办法，只能升内存，两根8G共16G的内存换上，结果还是一样，内存瞬间被秒杀（CPU利用率在0%徘徊）。由于是PC机，内存插槽共俩，目前市面上最大的单根内存为16G（价格1K+），就算买回来估计内存还是不够（卧槽，PC机伤不起啊），看样子别无它法 -- 删数据！！！

　　删除数据 - 说的容易， 不就是DELETE吗？靠，如果真这么干，我XXX估计能“知道上海凌晨4点的样子”（KB，Sorry，谁让我是XXX的Programmer，哥在这方面绝对比你牛X），而且估计会暴库（磁盘空间不足，产生的日志文件太大了）。

二、沙场点兵 - 众里寻他千百度

　　为了更好地阐述我所遇到的困难和问题，有必要做一些必要的测试和说明，同时这也是对如何解决问题的一种探究。因为毕竟这个问题的根本是如何来更好更快的操作数据，说到底就是DELETE、UPDATE、INSERT、TRUNCATE、DROP等的优化操作组合，我们的目的就是找出最优最快最好的方法。为了便于测试，准备了一张测试表Employee

<span>--</span><span>Create table Employee</span>
<span>CREATE</span> <span>TABLE</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span> (
    <span>[</span><span>EmployeeNo</span><span>]</span> <span>INT</span> <span>PRIMARY</span> <span>KEY</span>,
    <span>[</span><span>EmployeeName</span><span>]</span> <span>[</span><span>nvarchar</span><span>]</span>(<span><strong>50</strong></span>) <span>NULL</span>,
    <span>[</span><span>CreateUser</span><span>]</span> <span>[</span><span>nvarchar</span><span>]</span>(<span><strong>50</strong></span>) <span>NULL</span>,
    <span>[</span><span>CreateDatetime</span><span>]</span> <span>[</span><span>datetime</span><span>]</span> <span>NULL</span>
);

1. 数据插入PK

1.1. 循环插入，执行时间为38026毫秒

<span>--</span><span>循环插入</span>
<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Index</span> <span>INT</span> <span>=</span> <span><strong>1</strong></span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>WHILE</span> <span>@Index</span> <span> <span><strong>100000</strong></span>
<span>BEGIN</span>
    <span>INSERT</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>(EmployeeNo, EmployeeName, CreateUser, CreateDatetime) <span>VALUES</span>(<span>@Index</span>, <span>'</span><span>Employee_</span><span>'</span> <span>+</span> <span>CAST</span>(<span>@Index</span> <span>AS</span> <span>CHAR</span>(<span><strong>6</strong></span>)), <span>'</span><span>system</span><span>'</span>, <span>GETDATE</span>());
    <span>SET</span> <span>@Index</span> <span>=</span> <span>@Index</span> <span>+</span> <span><strong>1</strong></span>;
<span>END</span>

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;</span>

1.2.   事务循环插入，执行时间为6640毫秒

<span>--</span><span>事务循环</span>
<span>BEGIN</span> <span>TRAN</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Index</span> <span>INT</span> <span>=</span> <span><strong>1</strong></span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>WHILE</span> <span>@Index</span> <span> <span><strong>100000</strong></span>
<span>BEGIN</span>
    <span>INSERT</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>(EmployeeNo, EmployeeName, CreateUser, CreateDatetime) <span>VALUES</span>(<span>@Index</span>, <span>'</span><span>Employee_</span><span>'</span> <span>+</span> <span>CAST</span>(<span>@Index</span> <span>AS</span> <span>CHAR</span>(<span><strong>6</strong></span>)), <span>'</span><span>system</span><span>'</span>, <span>GETDATE</span>());
    <span>SET</span> <span>@Index</span> <span>=</span> <span>@Index</span> <span>+</span> <span><strong>1</strong></span>;
<span>END</span>

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

<span>COMMIT</span>;</span>

1.3.   批量插入，执行时间为220毫秒

<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>INSERT</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>(EmployeeNo, EmployeeName, CreateUser, CreateDatetime)
<span>SELECT</span> <span>TOP</span>(<span><strong>100000</strong></span>) EmployeeNo <span>=</span> ROW_NUMBER() <span>OVER</span> (<span>ORDER</span> <span>BY</span> C1.<span>[</span><span>OBJECT_ID</span><span>]</span>), <span>'</span><span>Employee_</span><span>'</span>, <span>'</span><span>system</span><span>'</span>, <span>GETDATE</span>()
<span>FROM</span> SYS.COLUMNS <span>AS</span> C1 <span>CROSS</span> <span>JOIN</span> SYS.COLUMNS <span>AS</span> C2
<span>ORDER</span> <span>BY</span> C1.<span>[</span><span>OBJECT_ID</span><span>]</span>

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

1.4.   CTE插入，执行时间也为220毫秒

<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

;<span>WITH</span> CTE(EmployeeNo, EmployeeName, CreateUser, CreateDatetime) <span>AS</span>(
    <span>SELECT</span> <span>TOP</span>(<span><strong>100000</strong></span>) EmployeeNo <span>=</span> ROW_NUMBER() <span>OVER</span> (<span>ORDER</span> <span>BY</span> C1.<span>[</span><span>OBJECT_ID</span><span>]</span>), <span>'</span><span>Employee_</span><span>'</span>, <span>'</span><span>system</span><span>'</span>, <span>GETDATE</span>()
    <span>FROM</span> SYS.COLUMNS <span>AS</span> C1 <span>CROSS</span> <span>JOIN</span> SYS.COLUMNS <span>AS</span> C2
    <span>ORDER</span> <span>BY</span> C1.<span>[</span><span>OBJECT_ID</span><span>]</span>
)
<span>INSERT</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span> <span>SELECT</span> EmployeeNo, EmployeeName, CreateUser, CreateDatetime <span>FROM</span> CTE;

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

小结：

• 按执行时间，效率依次为：CTE和批量插入效率相当，速度最快，事务插入次之，单循环插入速度最慢；
• 单循环插入速度最慢是由于INSERT每次都有日志，事务插入大大减少了写入日志次数，批量插入只有一次日志，CTE的基础是CLR，善用速度是最快的。

 

2.  数据删除PK

2.1.   循环删除，执行时间为1240毫秒

<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>DELETE</span> <span>FROM</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>;

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

2.2.  批量删除，执行时间为106毫秒

<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>SET</span> <span>ROWCOUNT</span> <span><strong>100000</strong></span>;

<span>WHILE</span> <span><strong>1</strong></span> <span>=</span> <span><strong>1</strong></span>
<span>BEGIN</span>
    <span>BEGIN</span> <span>TRAN</span>
    <span>DELETE</span> <span>FROM</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>;
    <span>COMMIT</span>
    <span>IF</span> <span><strong>@@ROWCOUNT</strong></span> <span>=</span> <span><strong>0</strong></span>
        <span>BREAK</span>;
<span>END</span>

<span>SET</span> <span>ROWCOUNT</span> <span><strong>0</strong></span>;

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

2.3.  TRUNCATE删除，执行时间为0毫秒

<span>SET</span> <span>STATISTICS</span> TIME <span>ON</span>;
<span>DECLARE</span> <span>@Timer</span> <span>DATETIME</span> <span>=</span> <span>GETDATE</span>();

<span>TRUNCATE</span> <span>TABLE</span> <span>[</span><span>dbo</span><span>]</span>.<span>[</span><span>Employee</span><span>]</span>;

<span>SELECT</span> <span>DATEDIFF</span>(MS, <span>@Timer</span>, <span>GETDATE</span>()) <span>AS</span> <span>[</span><span>执行时间(毫秒)</span><span>]</span>;
<span>SET</span> <span>STATISTICS</span> TIME <span>OFF</span>;

 小结：

• TRUNCATE太快了，清除10W数据一点没压力，批量删除次之，最后的DELTE太慢了；
• TRUNCATE快是因为它属于DDL语句，只会产生极少的日志，普通的DELETE不仅会产生日志，而且会锁记录。

 

三、磨刀霍霍 - 犹抱琵琶半遮面

　　由上面的第二点我们知道，插入最快和删除最快的方式分别是批量插入和TRUNCATE，所以为了达到删除大数据的目的，我们也将采用这两种方式的组合，其中心思想是先把需要保留的数据存放之新表中，然后TRUNCATE原表中的数据，最后再批量把数据插回去，当然实现方式也可以随便变通。

1. 保留需要的数据之新表中->TRUNCATE原表数据->还原之前保留的数据之原表中

　　脚本类似如下

<span>SELECT</span> <span>*</span> <span>INTO</span> #keep <span>FROM</span> Original <span>WHERE</span> CreateDate <span>&gt;</span> <span>'</span><span>2011-12-31</span><span>'</span>
<span>TRUNCATE</span> <span>TABLE</span> Original
<span>INSERT</span> Original <span>SELECT</span> <span>*</span> <span>FROM</span> #keep

　　第一条语句会把所有要保留的数据先存放至表#keep中（表#keep无需手工创建，由SELECT INTO生效），#keep会Copy原始表Original的表结构。PS：如果你只想创建表结构，但不拷贝数据，则对应的脚本如下

<span>SELECT</span> <span>*</span> <span>INTO</span> #keep <span>FROM</span> Original <span>WHERE</span> <span><strong>1</strong></span> <span>=</span> <span><strong>2</strong></span>

　　第二条语句用于清除整个表中数据，产生的日志文件基本可以忽略；第三条语句用于还原保留数据。

几点说明：

• 你可以不用SELECT INTO，自己通过写脚本（或拷贝现有表）来创建#keep，但是后者有一个弊端，即无法通过SQL脚本来获得对应的表生成Script（我的意思是和原有表完全一致的脚本，即基本列，属性，索引，约束等），而且当要操作的表比较多时，估计你肯定会抓狂；
• 既然第一点欠妥，那考虑新建一个同样的数据库怎么样？既可以使用现有脚本，而且生成的数据库基本一致，但是我告诉你最好别这么做，因为第一要跨库，第二，你得准备足够的磁盘空间。

 

2. 新建表结构->批量插入需要保留的数据->DROP原表->重命名新表为原表

　　CREATE TABLE #keep AS (xxx) xxx -- 使用上面提到的方法（使用既有表的创建脚本），但是不能够保证完全一致；

　　INSERT #keep SELECT * FROM Original where clause

　　DROP TBALE Original

　　EXEC SP_RENAME '#keep','Original'

　　这种方式比第一种方法略快点，因为省略了数据还原（即最后一步的数据恢复），但是稍微麻烦点，因为你需要创建一张和以前原有一模一样的表结构，包括基本列、属性、约束、索性等等。

三、数据收缩 - 秋风少落叶

 　　数据删除后，发现数据库占用空间大小并没有发生变化，此时我们就用借助强悍的数据收缩功能了，脚本如下，运行时间不定，取决于你的数据库大小，多则几十分钟，少则瞬间秒杀

<span>DBCC</span> SHRINKDATABASE(<span>DB_NAME</span>)