为什么窗口聚合函数的逻辑读取如此之高？

Question

我发现在使用公共子表达式假脱机的执行计划中，报告的逻辑读取对于大型表来说非常高。

经过反复试验，我找到了一个似乎适用于下面的测试脚本和执行计划的公式。 Worktable logical reads = 1 + NumberOfRows * 2 + NumberOfGroups * 4

我不明白为什么这个公式成立。这比我认为看计划的必要性要多。任何人都可以详细说明这是怎么回事吗？

或者如果失败了，是否有任何方法可以跟踪每次逻辑读取中读取的页面，以便我自己解决？

SET STATISTICS IO OFF; SET NOCOUNT ON;

IF Object_id('tempdb..#Orders') IS NOT NULL
  DROP TABLE #Orders;

CREATE TABLE #Orders
  (
     OrderID    INT IDENTITY(1, 1) NOT NULL PRIMARY KEY CLUSTERED,
     CustomerID NCHAR(5) NULL,
     Freight    MONEY NULL,
  );

CREATE NONCLUSTERED INDEX ix
  ON #Orders (CustomerID)
  INCLUDE (Freight);

INSERT INTO #Orders
VALUES (N'ALFKI', 29.46), 
       (N'ALFKI', 61.02), 
       (N'ALFKI', 23.94), 
       (N'ANATR', 39.92), 
       (N'ANTON', 22.00);

SELECT PredictedWorktableLogicalReads = 
        1 + 2 * Count(*) + 4 * Count(DISTINCT CustomerID)
FROM   #Orders;

SET STATISTICS IO ON;

SELECT OrderID,
       Freight,
       Avg(Freight) OVER (PARTITION BY CustomerID) AS Avg_Freight
FROM   #Orders; 

输出

PredictedWorktableLogicalReads
------------------------------
23

---

Table 'Worktable'. Scan count 3, logical reads 23, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.
Table '#Orders___________000000000002'. Scan count 1, logical reads 2, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

附加信息：

Query Tuning and Optimization Book 和 this blog post by Paul White 的第 3 章对这些线轴有很好的解释。

总之，计划顶部的段迭代器向它发送的行添加一个标志，指示它何时是新分区的开始。主段假脱机一次从段迭代器获取一行并将其插入到 tempdb 中的工作表中。一旦它得到表明一个新组已经开始的标志，它就会向嵌套循环运算符的顶部输入返回一行。这会导致在工作表中的行上调用流聚合，计算平均值，然后在截断工作表之前将该值与工作表中的行连接起来，以便为新组做好准备。段假脱机发出一个虚拟行，以便处理最后一组。

据我了解，工作表是一个堆（或者在计划中将其表示为索引假脱机）。但是，当我尝试复制相同的过程时，它只需要 11 次逻辑读取。

CREATE TABLE #WorkTable
  (
     OrderID    INT,
     CustomerID NCHAR(5) NULL,
     Freight    MONEY NULL,
  )

DECLARE @Average MONEY

PRINT 'Insert 3 Rows'

INSERT INTO #WorkTable
VALUES      (1, N'ALFKI', 29.46) /*Scan count 0, logical reads 1*/

INSERT INTO #WorkTable
VALUES      (2, N'ALFKI', 61.02) /*Scan count 0, logical reads 1*/

INSERT INTO #WorkTable
VALUES      (3, N'ALFKI', 23.94) /*Scan count 0, logical reads 1*/
PRINT 'Calculate AVG'

SELECT @Average = Avg(Freight)
FROM   #WorkTable /*Scan count 1, logical reads 1*/
PRINT 'Return Rows - With the average column included'

/*This convoluted query is just to force a nested loops plan*/
SELECT *
FROM   (SELECT @Average AS Avg_Freight) T /*Scan count 1, logical reads 1*/
       OUTER APPLY #WorkTable
WHERE  COALESCE(Freight, OrderID) IS NOT NULL
       AND @Average IS NOT NULL

PRINT 'Clear out work table'

TRUNCATE TABLE #WorkTable

PRINT 'Insert 1 Row'

INSERT INTO #WorkTable
VALUES      (4, N'ANATR', 39.92) /*Scan count 0, logical reads 1*/
PRINT 'Calculate AVG'

SELECT @Average = Avg(Freight)
FROM   #WorkTable /*Scan count 1, logical reads 1*/
PRINT 'Return Rows - With the average column included'

SELECT *
FROM   (SELECT @Average AS Avg_Freight) T /*Scan count 1, logical reads 1*/
       OUTER APPLY #WorkTable
WHERE  COALESCE(Freight, OrderID) IS NOT NULL
       AND @Average IS NOT NULL

PRINT 'Clear out work table'

TRUNCATE TABLE #WorkTable

PRINT 'Insert 1 Row'

INSERT INTO #WorkTable
VALUES      (5, N'ANTON', 22.00) /*Scan count 0, logical reads 1*/
PRINT 'Calculate AVG'

SELECT @Average = Avg(Freight)
FROM   #WorkTable /*Scan count 1, logical reads 1*/
PRINT 'Return Rows - With the average column included'

SELECT *
FROM   (SELECT @Average AS Avg_Freight) T /*Scan count 1, logical reads 1*/
       OUTER APPLY #WorkTable
WHERE  COALESCE(Freight, OrderID) IS NOT NULL
       AND @Average IS NOT NULL

PRINT 'Clear out work table'

TRUNCATE TABLE #WorkTable

PRINT 'Calculate AVG'

SELECT @Average = Avg(Freight)
FROM   #WorkTable /*Scan count 1, logical reads 0*/
PRINT 'Return Rows - With the average column included'

SELECT *
FROM   (SELECT @Average AS Avg_Freight) T
       OUTER APPLY #WorkTable
WHERE  COALESCE(Freight, OrderID) IS NOT NULL
       AND @Average IS NOT NULL

DROP TABLE #WorkTable 

Answer 1

工作表的逻辑读取计数不同：每个 行 读取有一个“逻辑读取”。这并不意味着工作台比“真正的”线轴工作台效率低（恰恰相反）；逻辑读取只是在不同的单元中。

我相信我的想法是计算工作表逻辑读取的散列页面不会很有用，因为这些结构是服务器内部的。报告逻辑读取计数器中假脱机的行使该数字对于分析目的更有意义。

这种洞察力应该使您的公式起作用的原因变得清晰。两个辅助线轴被完全读取两次 (2 * COUNT(*))，主线轴发出 (组值数 + 1) 行，如我的博客条目中所述，给出 (COUNT(DISTINCT CustomerID) + 1) 组件.加一用于主线轴发出的额外行，以指示最后一组已结束。

保罗

Answer 2

在公式中，您给出的 NumberOfRows * 2 将成立，因为执行图中的 Sort 函数和 Stream Aggregate 显示都需要所有行来完成处理。您能否确认添加“where”子句时逻辑读取的减少：

1. 运费价值
2. 客户 ID