本周技术研究部(TRD)的一名DBA 对我们编写SQL时的一些问题,进行了汇报讲演,以下是来自它的脚本,我在它讲演的基础上写出了自己想表述的,以便于大家相互交流学习。
/*--注意:准备数据(可略过,非常耗时)CREATE TABLE CHECK1_T1( ID INT, C1 CHAR(8000))CREATE TABLE CHECK1_T2( ID INT, C1 CHAR(8000))DECLARE @I INTSET @I=1WHILE @I<=10000 BEGIN INSERT INTO CHECK1_T1 SELECT @I,'C1' INSERT INTO CHECK1_T2 SELECT [email protected],'C1' SET @[email protected]+1 ENDCREATE TABLE CHECK2_T1( ID INT, C1 CHAR(8000))DECLARE @I INTSET @I=1WHILE @I<=10000 BEGIN INSERT INTO CHECK2_T1 SELECT @I,'C1' SET @[email protected]+1 ENDINSERT INTO CHECK2_T1 VALUES(10001,'C2')INSERT INTO CHECK2_T1 VALUES(10002,'C1')CREATE TABLE CHECK3_T1( ID INT, C1 CHAR(7000))CREATE TABLE CHECK3_T2( ID INT, C1 CHAR(7000))DECLARE @I INTSET @I=1WHILE @I<=20000 BEGIN IF @I%2 =0 BEGIN INSERT INTO CHECK3_T1 SELECT @I,'C1' END ELSE BEGIN INSERT INTO CHECK3_T1 SELECT @I,'C2' END IF @I%100=0 BEGIN INSERT INTO CHECK3_T2 SELECT @I,'C1' INSERT INTO CHECK3_T2 SELECT @I+50000,'C2' END SET @[email protected]+1 ENDCREATE TABLE CHECK4_T1( ID INT, C1 CHAR(500),)DECLARE @I INTSET @I=1WHILE @I<=500000 BEGIN IF @I%100000 =0 BEGIN INSERT INTO CHECK4_T1 SELECT @I,'C2' END ELSE BEGIN INSERT INTO CHECK4_T1 SELECT @I,'C1' END SET @[email protected]+1 ENDCREATE NONCLUSTERED INDEX NCIX_C1 ON CHECK4_T1(C1)CREATE TABLE CHECK5_T1( ID INT, C1 CHAR(10),)DECLARE @I INTSET @I=1WHILE @I<=10000 BEGIN INSERT INTO CHECK5_T1 SELECT @I,'C1' IF @I%2=0 BEGIN INSERT INTO CHECK5_T1 SELECT @I,'C1' END SET @[email protected]+1 END*/--=====================================--1、 Union all 代替 Union DBCC DROPCLEANBUFFERSDBCC FREEPROCCACHE --测试一:(26s) 执行计划:表扫描->排序->合并联接SELECT ID,C1 FROM CHECK1_T1 --1W条数据UNION SELECT ID,C1 FROM CHECK1_T2 --1W条数据--测试二: (4s) 执行计划:表扫描->表扫描串联SELECT ID,C1 FROM CHECK1_T1 --1W条数据UNION ALLSELECT ID,C1 FROM CHECK1_T2 --1W条数据--总结:测试一中的union 排序和去重合并是相当耗时的,如果不要此功能,大数据时最好加上ALL--=====================================--2、 Exists 代替 Count(*)DBCC DROPCLEANBUFFERSDBCC FREEPROCCACHE ----测试一: (7s) 执行计划:表扫描-> 流聚合-> 计算矢量 DECLARE @COUNT INT SELECT @COUNT=COUNT(*) FROM CHECK2_T1 WHERE C1='C1' --1W条数据 IF @COUNT>0 BEGIN PRINT 'S' END----测试二: (0s) 执行计划:常量扫描/表扫描-> 嵌套循环-> 计算标量 IF EXISTS(SELECT 1 FROM CHECK2_T1 WHERE C1='C1') --1W条数据 BEGIN PRINT 'S' END --总结:判断是否存在,用Exist即可,没必要用COUNT(*)将表的所有记录统计出来,扫描一次 --=====================================--3、 IN(Select COL1 From Table)的代替方式DBCC DROPCLEANBUFFERSDBCC FREEPROCCACHE --测试一: (3s)执行计划:表扫描 -> 哈希匹配 SELECT ID,C1 FROM CHECK3_T2 --400行WHERE ID IN (SELECT ID FROM CHECK3_T1 WHERE C1='C1') --2W行--测试二:(1s)执行计划:表扫描-> 并行度 -> 位图 -> 排序 -> 合并联接 -> 并行度SELECT A.ID,A.C1 FROM CHECK3_T2 A INNER JOIN CHECK3_T1 B ON A.ID=B.ID WHERE B.C1='C1' --测试三:(3s)执行计划:表扫描-> 哈希匹配 SELECT A.ID,A.C1 FROM CHECK3_T2 AWHERE EXISTS (SELECT 1 FROM CHECK3_T1 B WHERE B.ID=A.ID AND B.C1='C1')--总结:能用INNER JOIN 尽量用它,SQL SERVER在查询时会将关联表进行优化--=====================================--4、 Not Exists 代替 Not In--测试一:(8s) 执行计划:表扫描-> 嵌套循环 -> 哈希匹配SELECT ID,C1 FROM CHECK3_T1 --2W行WHERE ID NOT IN (SELECT ID FROM CHECK3_T2 WHERE C1='C1') --400行--测试二:(4s) 执行计划:表扫描-> 哈希匹配SELECT A.ID,A.C1 FROM CHECK3_T1 AWHERE NOT EXISTS (SELECT 1 FROM CHECK3_T2 B WHERE B.ID=A.ID AND B.C1='C1')--总结:尽量不使用NOT IN ,因为会调用嵌套循环,建议使用NOT EXISTS代替NOT IN--=====================================--5、 避免在条件列上使用任何函数DROP TABLE CHECK4_T1 CREATE NONCLUSTERED INDEX NCIX_C1 ON CHECK4_T1(C1) --加上非聚集索引---测试一:(4s)执行计划: 索引扫描SELECT * FROM CHECK4_T1 WHERE RTRIM(C1)='C2'---测试二:(0s)执行计划: 索引查找SELECT * FROM CHECK4_T1 WHERE C1='C2'--总结:where条件里对索引字段使用了函数,会使索引查找变成索引扫描,从而查询效率大幅下降--=====================================--6、 用sp_executesql执行动态sql DBCC DROPCLEANBUFFERSDBCC FREEPROCCACHE CREATE PROC UP_CHECK5_T1 ( @ID INT)AS SET NOCOUNT ON DECLARE @count INT, @sql NVARCHAR(4000) SET @sql = 'SELECT @count=count(*) FROM CHECK5_T1 WHERE ID = @ID' EXEC sp_executesql @sql, [email protected] INT OUTPUT, @ID int', @count OUTPUT, @ID PRINT @count CREATE PROC UP_CHECK5_T2 ( @ID INT)AS SET NOCOUNT ON DECLARE @sql NVARCHAR(4000) SET @sql = 'DECLARE @count INT;SELECT @count=count(*) FROM CHECK5_T1 WHERE ID = ' + CAST(@ID AS VARCHAR(10)) + ';PRINT @count' EXEC(@sql) ---测试一:瞬时DECLARE @N INTSET @N=1WHILE @N<=1000BEGIN EXEC UP_CHECK5_T1 @N SET @[email protected]+1END---测试二:2sDECLARE @N INTSET @N=1WHILE @N<=1000BEGIN EXEC UP_CHECK5_T2 @N SET @[email protected]+1ENDCREATE CLUSTERED INDEX CIX_ID ON CHECK5_T1(ID)DBCC DROPCLEANBUFFERSDBCC FREEPROCCACHE --查看缓存计划SELECT a.size_in_bytes '占用字节数', total_elapsed_time / execution_count '平均时间', total_logical_reads / execution_count '逻辑读', usecounts '重用次数', SUBSTRING(d.text, (statement_start_offset / 2) + 1, ((CASE statement_end_offset WHEN -1 THEN DATALENGTH(text) ELSE statement_end_offset END - statement_start_offset) / 2) + 1) '语句'FROM sys.dm_exec_cached_plans a CROSS apply sys.dm_exec_query_plan(a.plan_handle) c, sys.dm_exec_query_stats b CROSS apply sys.dm_exec_sql_text(b.sql_handle) dWHERE a.plan_handle = b.plan_handleORDER BY total_elapsed_time / execution_count DESC; --总结:通过执行下面缓存计划可以看出,第一种完全使用了缓存计划,查询达到了很好的效果;--而第二种则将缓存计划浪费了,导致缓存很快被占满,这种做法是相当不可取的--=====================================--7、 Left Join 的替代法--测试一 执行计划:表扫描 -> 哈希匹配SELECT A.ID,A.C1 FROM CHECK3_T1 A --2W行LEFT JOIN CHECK3_T2 B ON A.ID=B.ID WHERE B.C1='C1' --400行--测试二 执行计划:表扫描 -> 哈希匹配SELECT A.ID,A.C1 FROM CHECK3_T1 A RIGHT JOIN CHECK3_T2 B ON A.ID=B.ID WHERE a.C1='C1'--测试三 执行计划:表扫描 -> 哈希匹配SELECT A.ID,A.C1 FROM CHECK3_T1 A INNER JOIN CHECK3_T2 B ON A.ID=B.ID WHERE B.C1='C1'--总结:三条语句,在执行计划上完全一样,都是走的INNER JOIN的计划,--因为测试一和测试二中,WHERE语句都包含了LEFT 和RIGHT表的字段,SQLSERVER若发现只要有这个表的字段,则会自动按照INNER JOIN进行处理--补充测试:(1s)执行计划:表扫描-> 并行度 -> 位图 -> 排序 -> 合并联接 -> 并行度SELECT A.ID,A.C1 FROM CHECK3_T2 A --400行INNER JOIN CHECK3_T1 B ON A.ID=B.ID WHERE A.C1='C1' --2W行--总结:这里有一个比较有趣的地方,若主表和关联表数据差别很大时,走的执行计划走的另一条路--=====================================--8、 ON(a.id=b.id AND a.tag=3)--测试一SELECT A.ID,A.C1 FROM CHECK3_T1 A INNER JOIN CHECK3_T2 B ON A.ID=B.ID AND A.C1='C1'--测试二SELECT A.ID,A.C1 FROM CHECK3_T1 A INNER JOIN CHECK3_T2 B ON A.ID=B.ID WHERE A.C1='C1'--总结:内连接:无论是左表和右表的筛选条件都可以放到WHERE子句中--测试一SELECT A.ID,A.C1,B.C1 FROM CHECK3_T1 A LEFT JOIN CHECK3_T2 B ON A.ID=B.ID AND B.C1='C1'--测试二SELECT A.ID,A.C1,B.C1 FROM CHECK3_T1 A LEFT JOIN CHECK3_T2 B ON A.ID=B.ID WHERE B.C1='C1'--总结:左外连接:当右表中的过滤条件放入ON子句后和WHERE子句后的结果不一样--=====================================--9、 赋值给变量,加Top 1--测试一:(3s) 执行计划:表扫描DECLARE @ID INTSELECT @ID=ID FROM CHECK1_T1 WHERE C1='C1'SELECT @ID --测试二:(0s)执行计划:表扫描-> 前几行DECLARE @ID INTSELECT TOP 1 @ID=ID FROM CHECK1_T1 WHERE C1='C1'SELECT @ID--总结:给变量赋值最好都加上TOP 1,一从查询效率上增强,二为了准确性,若表CHECK1_T1有多个值,[email protected]--=====================================--10、 考虑是否适合用CASE语句DECLARE @S INT=1SELECT * FROM CHECK5_T1WHERE C1=(CASE @S WHEN 1 THEN C1 ELSE 'C2' END)SELECT * FROM CHECK5_T1WHERE @S=1 OR C1='C2'/*--=====================================12、检查语句是否需要Distinct. 执行计划:表扫描-> 哈希匹配-> 并行度-> 排序select distinct c1 from CHECK3_T1 13、禁用Select *,指定具体列名select c1 from CHECK4_T1select * from CHECK4_T114、Insert into Table(*),指定具体的列名15、Isnull,没有必要的时候不要对字段使用isnull,同样会产生无法有效利用索引的问题, 和避免在筛选列上使用函数同样的原理。 16、嵌套子查询,加上查询条件,确保子查询的结果集最小--=====================================*/