SORT 操作的瓶颈

Question

我有以下查询，事实证明它的成本非常高，需要 6-8 秒才能执行。查看执行计划，SORT 操作的成本为 79%。我可以在这里得到任何改进吗？

    SELECT
        A.StageName,
        C.Month,
        C.MonthName as Label,
        C.Year,
        isnull(A.Average,0) as Data
    FROM
    (   
        SELECT 
            S.StageName,
            MONTH(TimeIn) as MonthNumber, 
            DATENAME(MONTH,TimeIn) as Month, 
            YEAR(TimeIn) as Year, 
            ISNULL(AVG(DATEDIFF(mi,TimeIn,isnull(TimeOut,@TodayDate))),0) as Average
        FROM
            VisitMovement VM
        INNER JOIN Stage S on
            VM.StageID = S.StageID
        WHERE 
            (VM.TimeIn >= @StartDate AND 
            VM.TimeIn  < DATEADD (d,1,@EndDate)) AND
            (VM.TimeOut < DATEADD (d,1,@EndDate) OR VM.TimeOut IS NULL)
        GROUP BY
            S.StageNumber,
            S.StageName,
            MONTH(TimeIn), 
            DATENAME(MONTH,TimeIn), 
            YEAR(TimeIn)
    ) A
    RIGHT JOIN (select distinct Month,MonthName,Year from Calendar WHERE DATE >= @StartDate AND DATE < DATEADD (d,1,@EndDate)) C on
        A.MonthNumber = C.Month and
        A.Month = C.MonthName and
        A.Year = C.Year 
    GROUP BY
        A.StageName,
        C.Month,
        C.MonthName,
        C.Year,
        A.Average   
    ORDER BY  
        CASE WHEN @Ordering = 'asc'  THEN C.Year            END ASC, 
        CASE WHEN @Ordering = 'asc'  THEN C.Month           END ASC,
        CASE WHEN @Ordering = 'asc'  THEN A.StageName       END ASC,
        CASE WHEN @Ordering = 'desc' THEN C.Year            END DESC, 
        CASE WHEN @Ordering = 'desc' THEN C.Month           END DESC,
        CASE WHEN @Ordering = 'desc'  THEN A.StageName      END DESC

Answer 1

因为 order by 必须评估每一行，我认为它不能优化使用索引。将 order by 替换为 row_number() 进行默认排序，并且仅反转一次顺序应该至少可以防止对单行的 @Ordering 进行多次评估。

在下面的伪代码中，原始查询放在 CTE 中。 Row_number 确定升序排序，在 cte 以下，如果需要则颠倒顺序：

;with cte as
(
    SELECT
        A.StageName,
        C.Month,
        ...,
        row_number() over (order by C.Year,C.Month,A.StageName) sortOrder
    FROM
       ...rest of the query, excluding the order by
)
select * --or list the columns without the sortOrder 
from cte
order by sortOrder * case @Ordering when 'desc' then -1 else 1 end

Answer 2

虽然我知道你不能去掉子查询中不同列上的GROUP BY，但你可以让系统更容易。

目前你有

S.StageNumber,
S.StageName,
MONTH(TimeIn), 
DATENAME(MONTH,TimeIn), 
YEAR(TimeIn)

我想要处理的数据相当多。请允许我做一些猜测：

S.StageNumber,  -- int, 4 bytes
S.StageName, -- string, 20 bytes
MONTH(TimeIn),  -- int, 4 bytes
DATENAME(MONTH,TimeIn),  -- string 5 bytes
YEAR(TimeIn) -- int, 4 byte

现在，有一些依赖项：

如果您知道 MONTH（数字），那么您现在也知道它的名称
我猜 StageName + StageNumber 是唯一的并且与 StageID 直接相关。如果不是，您可能需要在外层再次 GROUP BY。

那会把我们带到

S.StageID,  -- int, 4 bytes
MONTH(TimeIn),  -- int, 4 bytes
YEAR(TimeIn) -- int, 4 byte

这意味着 GROUP BY 的排序必须运行通过每条记录仅 12 个字节而不是以前的每条记录 37 个字节，并且数字排序比字符串快很多无论如何（例如，由于 upper/lower 大小写、重音等）

我已尝试相应地重写查询（未经测试！）。我还将月份信息的获取移动到一个单独的临时 table，这应该对查询优化器有一点帮助。

SELECT DISTINCT Month,MonthName,Year 
  INTO #dates
  FROM Calendar 
 WHERE DATE >= @StartDate AND DATE < DATEADD (d,1,@EndDate)

CREATE UNIQUE CLUSTERED INDEX uq0_#dates ON #dates (Month,Year)

SELECT
        A.StageName,
        C.Month,
        C.MonthName as Label,
        C.Year,
        isnull(A.Average,0) as Data
    FROM 
    (   
        SELECT S.StageName,
               MonthNumber,
               Year,
               Average
          FROM (    
                  SELECT VM.StageID,
                         MONTH(TimeIn) as MonthNumber, 
                         YEAR(TimeIn) as Year, 
                         ISNULL(AVG(DATEDIFF(mi,TimeIn,isnull(TimeOut,@TodayDate))),0) as Average
                    FROM
                         VisitMovement VM
                   WHERE 
                        (VM.TimeIn >= @StartDate AND 
                        VM.TimeIn  < DATEADD (d,1,@EndDate)) AND
                        (VM.TimeOut < DATEADD (d,1,@EndDate) OR VM.TimeOut IS NULL)
                  GROUP BY
                        VM.StageID,
                        MONTH(TimeIn), 
                        YEAR(TimeIn)
               ) grp
          JOIN Stage S 
            ON S.StageID = grp.StageID
        )
    ) A
    RIGHT OUTER JOIN #dates C
    on
        A.MonthNumber = C.Month and
        -- A.Month = C.MonthName and
        A.Year = C.Year 
    ORDER BY  
        CASE WHEN @Ordering = 'asc'  THEN C.Year            END ASC, 
        CASE WHEN @Ordering = 'asc'  THEN C.Month           END ASC,
        CASE WHEN @Ordering = 'asc'  THEN A.StageName       END ASC,
        CASE WHEN @Ordering = 'desc' THEN C.Year            END DESC, 
        CASE WHEN @Ordering = 'desc' THEN C.Month           END DESC,
        CASE WHEN @Ordering = 'desc'  THEN A.StageName      END DESC

希望对您有所帮助。

SORT 操作的瓶颈

Bottle neck on SORT operation

sql

sql-server

query-performance

sql-server-2016