高性能空间连接 - 使用 SQL 地理函数的最近记录

Question

在 SQL Server 2012 中改进最接近的空间匹配性能（地理 STDistance）时，我发现迭代地增加搜索半径可以提高我们数据集的性能。

我有以下三步查询，我试图将其转换为最大深度为 3 的递归 CTE，并进行距离比较 [LEVEL]*500。

SELECT
    N.WorkID,
    M.WorkID AS MatchWorkID,
    N.Location.STDistance(M.Location) AS Meters
INTO 
    #MatchWork500
FROM 
    NewWork N
CROSS APPLY 
    (SELECT TOP (1) 
         C.WorkID, C.Location
     FROM 
         CurrWork C
     WHERE 
         N.Location.STDistance(C.Location) <= 500
     ORDER BY 
         N.Location.STDistance(C.Location)) AS M
SELECT
    N.WorkID,
    M.WorkID AS MatchWorkID,
    N.Location.STDistance(M.Location) AS Meters
INTO 
    #MatchWork1000
FROM 
    NewWork N
LEFT JOIN 
    #MatchWork500 M500 ON M500.WorkID = N.WorkID
CROSS APPLY 
    (SELECT TOP (1) 
         C.WorkID, C.Location
     FROM 
         CurrWork C
     WHERE 
         N.Location.STDistance(C.Location) <= 1000
     ORDER BY 
         N.Location.STDistance(C.Location)) AS M
WHERE 
    M500.WorkID IS NULL

SELECT
    N.WorkID,
    M.WorkID AS MatchWorkID,
    N.Location.STDistance(M.Location) AS Meters
INTO 
    #MatchWork1000
FROM 
    NewWork N
LEFT JOIN 
    #MatchWork500 M500 ON M500.WorkID = N.WorkID
LEFT JOIN 
    #MatchWork1000 M1000 ON M1000.WorkID = N.WorkID
CROSS APPLY 
    (SELECT TOP (1) 
         C.WorkID, C.Location
     FROM 
         CurrWork C
     WHERE 
         N.Location.STDistance(C.Location) <= 1500
     ORDER BY
         N.Location.STDistance(C.Location)) AS M
WHERE 
    M500.WorkID IS NULL 
    AND M1000.WorkID IS NULL

每次迭代只需要处理之前未匹配的内容，最好是单个 CTE table。请不要存储过程的答案，感谢帮助。

Answer 1

设法为此找到一个高性能的解决方案（剧透，我找不到一个高性能的 CTE 解决方案）。第一步是改进 CurrWork 上的空间索引。有问题的数据集都有 Geography 列。投影为WGS1984，供参考。这对于大多数基于英国的空间数据操作非常有效，尤其是对于 building/address 个位置；

    CREATE SPATIAL INDEX [IDX_Location] ON [CurrWork] ([Location])
    USING  GEOGRAPHY_GRID 
    WITH (
        GRIDS =(LEVEL_1 = HIGH,LEVEL_2 = MEDIUM,LEVEL_3 = MEDIUM,LEVEL_4 = MEDIUM), 
        CELLS_PER_OBJECT = 3, PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON
    ) ON [PRIMARY]
    GO

我发现在我的案例中我必须使用索引提示，这可能并不总是需要的。

    FROM CurrWork AS C WITH(INDEX([IDX_Location]))

为了找到最接近的 CurrWork，我发现排名距离作为第一步效果很好；

    RANK() OVER(PARTITION BY N.WorkID ORDER BY C.Location.STDistance(N.Location),C.WorkID) rnk

然后M.rnk = 1在连接条件中限制为仅最近的记录。完整代码如下...希望有人觉得它有用！

    DECLARE @rad FLOAT = 10000 --10km search radius

    SELECT 
    N.WorkID,
    M.MatchWorkID,
    M.Meters

    FROM NewWork N
    LEFT OUTER JOIN ( 
        SELECT 
        N.WorkID,
        C.WorkID AS MatchWorkID,
        C.Location.STDistance(N.Location) AS Meters,
        RANK() OVER(PARTITION BY N.WorkID ORDER BY C.Location.STDistance(N.Location),C.WorkID) rnk
        FROM CurrWork AS C WITH(INDEX([IDX_Location]))
        JOIN NewWork N ON C.Location.STDistance(N.Location) < @rad
    ) M ON N.WorkID = M.WorkID AND M.rnk = 1

高性能空间连接 - 使用 SQL 地理函数的最近记录

Performant spatial join - closest record using SQL Geography functions

sql-server

recursion

spatial

common-table-expression

sql-server-2012