索引搜索操作员成本

Pau*_*ite 10 sql-server execution-plan database-internals

对于下面的AdventureWorks示例数据库查询:

SELECT 
    P.ProductID, 
    CA.TransactionID
FROM Production.Product AS P
CROSS APPLY
(
    SELECT TOP (1)
        TH.TransactionID
    FROM Production.TransactionHistory AS TH
    WHERE
        TH.ProductID = P.ProductID
    ORDER BY 
        TH.TransactionID DESC
) AS CA;
Run Code Online (Sandbox Code Playgroud)

执行计划显示索引搜索估计操作员成本0.0850383 (93%) :

计划

成本与使用的基数估计模型无关。

它不是简单地将Estimated CPU CostEstimated I/O Cost相加。它也不是指数搜索的一次执行成本乘以估计执行次数

这个成本数字是如何得出的?

Pau*_*ite 12

完整的成本推导逻辑很复杂,但对于问题中相对简单的情况:

输入

  1. 操作符被执行的次数
    这是估计的执行次数504

  2. 索引中的基数(总行数)Index Seek运算符
    TableCardinality属性给出:113,443

  3. 索引中的数据页数:201
    这个数可以通过多种方式获得,例如从sys.allocation_units

    SELECT 
        AU.data_pages
    FROM sys.allocation_units AS AU
    JOIN sys.partitions AS P
        ON P.hobt_id = AU.container_id
    WHERE
        AU.[type_desc] = N'IN_ROW_DATA'
        AND P.[object_id] = OBJECT_ID(N'Production.TransactionHistory', N'U')
        AND P.index_id = 
            INDEXPROPERTY(P.[object_id], N'IX_TransactionHistory_ProductID', 'IndexID');
    
    Run Code Online (Sandbox Code Playgroud)
  4. 索引的密度(1 /不同值):0.002267574
    这在索引统计的密度向量中可用:

    DBCC SHOW_STATISTICS 
    (
        N'Production.TransactionHistory', 
        N'IX_TransactionHistory_ProductID'
    ) 
    WITH DENSITY_VECTOR;
    
    Run Code Online (Sandbox Code Playgroud)

    密度

计算

-- Input numbers
DECLARE
    @Executions float = 504,
    @Density float = 0.002267574,
    @IndexDataPages float = 201,
    @Cardinality float = 113443;

-- SQL Server cost model constants
DECLARE
    @SeqIO float = 0.000740740740741,
    @RandomIO float = 0.003125,
    @CPUbase float = 0.000157,
    @CPUrow float = 0.0000011;

-- Computation
DECLARE
    @IndexPages float = CEILING(@IndexDataPages * @Density),
    @Rows float = @Cardinality * @Density,
    @Rebinds float = @Executions - 1e0;

DECLARE
    @CPU float = @CPUbase + (@Rows * @CPUrow),
    @IO float = @RandomIO + (@SeqIO * (@IndexPages - 1e0)),
    -- sample with replacement
    @PSWR float = @IndexDataPages * (1e0 - POWER(1e0 - (1e0 / @IndexDataPages), @Rebinds));

-- Cost components (no rewinds)
DECLARE
    @InitialCost float = @RandomIO + @CPUbase + @CPUrow,
    @RebindCPU float = @Rebinds * (1e0 * @CPUbase + @CPUrow),
    @RebindIO float = (1e0 / @Rows) * ((@PSWR - 1e0) * @IO);

-- Result
SELECT 
    OpCost = @InitialCost + @RebindCPU + @RebindIO;
Run Code Online (Sandbox Code Playgroud)

数据库<>小提琴

结果

  • 据我所知,这方面的成本自 2005 年以来就没有改变。 (2认同)