统计--VARCHAR与NVARCHAR在统计预估上的区别

最近遇到一个问题，当查询使用到模糊查询时，由于预估返回行数过高，执行计划认为索引查找+Key Lookup的成本过高，因此采用Clustered Index Scan的方式，消耗大量逻辑IO,执行计划较差。

经过测试，发现对于模糊查询，NVARCHAR和VARCHAR的预估返回行数差距很大，因此拿出来供大家一起测试。

首先生成测试数据，分别创建TB101和TB102的表，表上有相同的聚集索引和非聚集索引，表中有100w数据，创建测试数据脚本如下:

DROP TABLE TB101
GO
DROP TABLE TB102
GO
SELECT
CAST(NCHAR(19968+20902*RAND(RID))+NCHAR(19968+20902/2*RAND(RID))+NCHAR(19968+20902/3*RAND(RID)) AS varchar(40)) AS RName
,*
INTO TB101
FROM(
SELECT ROW_NUMBER()OVER(ORDER BY T1.OBJECT_ID DESC) AS RID,T1.* FROM sys.all_objects T2
CROSS JOIN sys.all_columns T1
) AS T
WHERE T.RID<1000000
GO
SELECT * INTO TB102 FROM TB101
GO
ALTER TABLE TB102
ALTER COLUMN RName NVARCHAR(40)
GO
CREATE UNIQUE CLUSTERED INDEX IDX_RID ON TB101(RID)
GO
CREATE UNIQUE CLUSTERED INDEX IDX_RID ON TB102(RID)
GO
CREATE INDEX IDX_Name ON TB101(RName)
GO
CREATE INDEX IDX_Name ON TB102(RName)
GO

EXEC sp_spaceused ‘TB101‘
EXEC sp_spaceused ‘TB102‘
GO

SELECT * FROM TB101
WHERE RName LIKE ‘你好%‘

SELECT * FROM TB102
WHERE RName LIKE N‘你好%‘