还有3种方法，其中UNPIVOT(1)是目前为止最快的，其次是模拟UNPIVOT(3)，它比(1)慢得多，但仍然比(2)快得多

CREATE TABLE dates
    (
      number INT PRIMARY KEY ,
      date1 DATETIME ,
      date2 DATETIME ,
      date3 DATETIME ,
      cost INT
    )

INSERT  INTO dates
VALUES  ( 1, '1/1/2008', '2/4/2008', '3/1/2008', 10 )
INSERT  INTO dates
VALUES  ( 2, '1/2/2008', '2/3/2008', '3/3/2008', 20 )
INSERT  INTO dates
VALUES  ( 3, '1/3/2008', '2/2/2008', '3/2/2008', 30 )
INSERT  INTO dates
VALUES  ( 4, '1/4/2008', '2/1/2008', '3/4/2008', 40 )
GO

解决方案1 (UNPIVOT)

SELECT  number ,
        MAX(dDate) maxDate ,
        cost
FROM    dates UNPIVOT ( dDate FOR nDate IN ( Date1, Date2,
                                            Date3 ) ) as u
GROUP BY number ,
        cost 
GO

解决方案2(每行子查询)

SELECT  number ,
        ( SELECT    MAX(dDate) maxDate
          FROM      ( SELECT    d.date1 AS dDate
                      UNION
                      SELECT    d.date2
                      UNION
                      SELECT    d.date3
                    ) a
        ) MaxDate ,
        Cost
FROM    dates d
GO

方案3(模拟UNPIVOT)

;WITH    maxD
          AS ( SELECT   number ,
                        MAX(CASE rn
                              WHEN 1 THEN Date1
                              WHEN 2 THEN date2
                              ELSE date3
                            END) AS maxDate
               FROM     dates a
                        CROSS JOIN ( SELECT 1 AS rn
                                     UNION
                                     SELECT 2
                                     UNION
                                     SELECT 3
                                   ) b
               GROUP BY Number
             )
    SELECT  dates.number ,
            maxD.maxDate ,
            dates.cost
    FROM    dates
            INNER JOIN MaxD ON dates.number = maxD.number
GO

DROP TABLE dates
GO

我更喜欢基于case-when的解决方案，我的假设是，与其他可能的解决方案(如交叉应用、values()、自定义函数等)相比，它对可能的性能下降的影响应该最小。

下面是case-when版本，它在大多数可能的测试用例中处理空值:

SELECT
    CASE 
        WHEN Date1 > coalesce(Date2,'0001-01-01') AND Date1 > coalesce(Date3,'0001-01-01') THEN Date1 
        WHEN Date2 > coalesce(Date3,'0001-01-01') THEN Date2 
        ELSE Date3
    END AS MostRecentDate
    , *
from 
(values
     (  1, cast('2001-01-01' as Date), cast('2002-01-01' as Date), cast('2003-01-01' as Date))
    ,(  2, cast('2001-01-01' as Date), cast('2003-01-01' as Date), cast('2002-01-01' as Date))
    ,(  3, cast('2002-01-01' as Date), cast('2001-01-01' as Date), cast('2003-01-01' as Date))
    ,(  4, cast('2002-01-01' as Date), cast('2003-01-01' as Date), cast('2001-01-01' as Date))
    ,(  5, cast('2003-01-01' as Date), cast('2001-01-01' as Date), cast('2002-01-01' as Date))
    ,(  6, cast('2003-01-01' as Date), cast('2002-01-01' as Date), cast('2001-01-01' as Date))
    ,( 11, cast(NULL         as Date), cast('2002-01-01' as Date), cast('2003-01-01' as Date))
    ,( 12, cast(NULL         as Date), cast('2003-01-01' as Date), cast('2002-01-01' as Date))
    ,( 13, cast('2003-01-01' as Date), cast(NULL         as Date), cast('2002-01-01' as Date))
    ,( 14, cast('2002-01-01' as Date), cast(NULL         as Date), cast('2003-01-01' as Date))
    ,( 15, cast('2003-01-01' as Date), cast('2002-01-01' as Date), cast(NULL         as Date))
    ,( 16, cast('2002-01-01' as Date), cast('2003-01-01' as Date), cast(NULL         as Date))
    ,( 21, cast('2003-01-01' as Date), cast(NULL         as Date), cast(NULL         as Date))
    ,( 22, cast(NULL         as Date), cast('2003-01-01' as Date), cast(NULL         as Date))
    ,( 23, cast(NULL         as Date), cast(NULL         as Date), cast('2003-01-01' as Date))
    ,( 31, cast(NULL         as Date), cast(NULL         as Date), cast(NULL         as Date))

) as demoValues(id, Date1,Date2,Date3)
order by id
;

结果是:

MostRecent    id   Date1      Date2      Date3
2003-01-01    1    2001-01-01 2002-01-01 2003-01-01
2003-01-01    2    2001-01-01 2003-01-01 2002-01-01
2003-01-01    3    2002-01-01 2001-01-01 2002-01-01
2003-01-01    4    2002-01-01 2003-01-01 2001-01-01
2003-01-01    5    2003-01-01 2001-01-01 2002-01-01
2003-01-01    6    2003-01-01 2002-01-01 2001-01-01
2003-01-01    11   NULL       2002-01-01 2003-01-01
2003-01-01    12   NULL       2003-01-01 2002-01-01
2003-01-01    13   2003-01-01 NULL       2002-01-01
2003-01-01    14   2002-01-01 NULL       2003-01-01
2003-01-01    15   2003-01-01 2002-01-01 NULL
2003-01-01    16   2002-01-01 2003-01-01 NULL
2003-01-01    21   2003-01-01 NULL       NULL
2003-01-01    22   NULL       2003-01-01 NULL
2003-01-01    23   NULL       NULL       2003-01-01
NULL          31   NULL       NULL       NULL

从SQL Server 2012我们可以使用IIF。

 DECLARE @Date1 DATE='2014-07-03';
 DECLARE @Date2 DATE='2014-07-04';
 DECLARE @Date3 DATE='2014-07-05';

 SELECT IIF(@Date1>@Date2,
        IIF(@Date1>@Date3,@Date1,@Date3),
        IIF(@Date2>@Date3,@Date2,@Date3)) AS MostRecentDate

最后，针对以下内容:

SQL Server 2022 (16.x)预览 Azure SQL数据库 Azure SQL托管实例

我们也可以用GREATEST。与其他T-SQL函数类似，这里有一些重要的注意事项:

如果所有参数都具有相同的数据类型，并且 的类型为 支持 进行比较， GREATEST将返回该类型; 否则， 函数 将隐式地将所有参数转换为 最高优先级 的数据类型进行比较，并使用 这个 类型 作为返回类型; 如果一个或多个参数不是NULL，那么NULL参数将在比较过程中被忽略;如果所有参数都为NULL，则GREATEST将返回NULL;

在GREATEST中不支持以下类型进行比较:varchar(max)、varbinary(max)或nvarchar(max)超过8,000字节、游标、几何、地理、图像、非字节顺序的用户定义类型、ntext、table、text和xml。

下面是使用T-SQL和SQL Server实现Max功能的另一个很好的解决方案

SELECT [Other Fields],
  (SELECT Max(v) 
   FROM (VALUES (date1), (date2), (date3),...) AS value(v)) as [MaxDate]
FROM [YourTableName]

Values是表值构造函数。

"指定要构造到表中的一组行值表达式。Transact-SQL表值构造函数允许在单个DML语句中指定多行数据。表值构造函数可以指定为INSERT…VALUES语句，或者作为MERGE语句的USING子句或FROM子句中的派生表。

使用CROSS APPLY (2005+) ....

SELECT MostRecentDate 
FROM SourceTable
    CROSS APPLY (SELECT MAX(d) MostRecentDate FROM (VALUES (Date1), (Date2), (Date3)) AS a(d)) md