WITH 子句中的 Getdate() 是否应该声明为变量？

Question

我了解通常将当前日期声明为变量是有用的：

DECLARE @CurrentDate as DateTime
SET @CurrentDate = Getdate()

我的理解是，这会对当前日期/时间进行一次采样并将其记录为静态变量。这有助于确保“当前时间”在查询期间保持不变，并且还可以避免重复使用比获取变量值更昂贵的 Getdate()。

我的问题是....在递归 WITH 子句的情况下是否需要这样做，该子句似乎是一次获取日期？

例如，考虑以下代码：

WITH CalendarSequence as(
  SELECT Getdate() as RollingDate
  UNION ALL
  SELECT DateAdd(month, -1, RollingDate) as RollingDate
  FROM CalendarSequence
  WHERE DateAdd(month, -1, RollingDate) > Convert(date, '2016-01-01')
)
SELECT 
Year(CalendarSequence.RollingDate)*100+Month(CalendarSequence.RollingDate) as MissingYearMonth
FROM CalendarSequence
LEFT OUTER JOIN TableName
ON Year(CalendarSequence.RollingDate)*100+Month(CalendarSequence.RollingDate) = TableName.YearMonthField
WHERE TableName.YearMonthField is NULL

此查询生成一个临时日期表，并将其与任意表进行比较，以突出显示没有活动/数据的各个月份。

就这段特定的代码而言，Getdate() 是否被多次使用并不重要（除非在查询执行期间月份发生了变化！）但对于类似的查询而言，这并非普遍适用。我提出这个问题的部分动机是为了更好地了解 WITH 函数正在做什么以及它是否有任何令人惊讶的行为。

Answer 1

我在 SQL Server 2016 中运行了这两种方法。我没有发现任何区别。但是，正如 @Cato 在 cmets 中提到的，最好声明为变量并传递变量，而不是等待优化器处理它。

方法一

DECLARE @currentDate DATE = GETDATE()
;WITH CalendarSequence as(
  SELECT @currentDate as RollingDate
  UNION ALL
  SELECT DateAdd(month, -1, RollingDate) as RollingDate
  FROM CalendarSequence
  WHERE DateAdd(month, -1, RollingDate) > Convert(date, '2018-01-01')
)
SELECT * FROM CalendarSequence

方法2

;WITH CalendarSequence as(
  SELECT Getdate() as RollingDate
  UNION ALL
  SELECT DateAdd(month, -1, RollingDate) as RollingDate
  FROM CalendarSequence
  WHERE DateAdd(month, -1, RollingDate) > Convert(date, '2018-01-01')
)
SELECT * FROM CalendarSequence

它们都有相同的执行计划。完全没有变化。

Answer 2

执行摘要：如果您想要一个日期/时间的值，请将其捕获在一个变量中并根据需要使用它。它使您的意图清晰，并避免软件更新和其他细微之处可能出现的问题。

以下代码使用 SQL Server 2008 R2 进行测试，证明并非所有 select 语句都相同。虽然GetDate() 似乎是每个实例（也可能是所有实例）的runtime constant，但问题比人们想象的要微妙。

GetDate() 可以在所有列和行中保持不变。

-- Constant across all columns and rows.
with Murphy as (
  select GetDate() as A, GetDate() as B, 1 as Rows
  union all
  select GetDate(), GetDate(), Rows + 1
    from Murphy
    where A = B and Rows < 1000000 )
  select Min( A ) as MinA, Max( A ) as MaxA, Min( B ) as MinB, Max( B ) as MaxB
    from Murphy
    option ( MaxRecursion 0 );

更高效的版本使用cross join 而不是递归。

-- Constant across all columns and rows.
declare @Limit as Int = 100000;
with Ten ( Number ) as
  ( select * from ( values (0), (1), (2), (3), (4), (5), (6), (7), (8), (9) ) as Digits( Number ) ),
  TenUp2 ( Number ) as ( select 42 from Ten as L cross join Ten as R ),
  TenUp4 ( Number ) as ( select 42 from TenUp2 as L cross join TenUp2 as R ),
  TenUp8 ( Number ) as ( select 42 from TenUp4 as L cross join TenUp4 as R ),
  Numbers ( Number, A, B ) as ( select top (@Limit) Row_Number() over ( order by ( select NULL ) ),
    GetDate(), GetDate() from TenUp8 )
  select Min( A ) as MinA, Max( A ) as MaxA, Min( B ) as MinB, Max( B ) as MaxB
    from Numbers;

再一次，也许不同的实例会分开。

-- Fails (and does not generate an execution plan).
declare @A as DateTime = GetDate();
declare @B as DateTime = @A;
declare @Trials as Int = 0;

while @A = @B
  begin
  select @A = GetDate(), @B = GetDate(), @Trials += 1;
  if @Trials % 1000 = 0
    print @Trials;
  end

select @A as A, @B as B, @Trials as Trials;

所以你认为这只是一个冒充select 的set 和一个真正的 select，生成执行计划的工作方式会有所不同。

-- Fails.
declare @A as DateTime = GetDate();
declare @B as DateTime = @A;
declare @Trials as Int = 0;

while @A = @B
  begin
  select @A = GetDate(), @B = GetDate(), @Trials += 1
    from ( values ( 42 ) ) as PH( A );
  if @Trials % 1000 = 0
    print @Trials;
  end

select @A as A, @B as B, @Trials as Trials;

如果值来自表值构造函数怎么办？

-- Fails.
declare @A as DateTime = GetDate();
declare @B as DateTime = @A;
declare @Trials as Int = 0;

select @A = GetDate(), @B = @A;
while @A = @B
  begin
  select @A = A, @B = B, @Trials += 1
    from ( values ( GetDate(), GetDate() ) ) as PH( A, B )
  if @Trials % 1000 = 0
    print @Trials;
  end

select @A as A, @B as B, @Trials as Trials;

嗯，所有的失败都来自为变量赋值的select 语句。让我们通过向表中插入行并稍后检查它们来消除这种情况。 （注意：此示例在带有 SQL Server 2017 的 SQL Fiddle 上运行时不会失败。）

-- Fails on SQL Server 2008 R2, but not on SQL Server 2017.
declare @Samples as Table ( A DateTime, B DateTime );
declare @Trials as Int = 0;
while @Trials < 100000
  begin
  insert into @Samples ( A, B ) values ( GetDate(), GetDate() )
  set @Trials += 1;
  end

select A, B
  from @Samples
  where A != B;
select Min( A ) as MinA, Max( A ) as MaxA, Min( B ) as MinB, Max( B ) as MaxB
  from @Samples;