SQL UPDATE 查询 - 值取决于另一行

Question

There is a SQL Server database temporary table, let it be TableA. And the table structure is following:

CREATE TABLE #TableA 
( 
  ID BIGINT IDENTITY (1, 1) PRIMARY KEY,
  MapVal1 BIGINT NOT NULL,
  MapVal2 BIGINT NOT NULL,
  IsActual BIT NULL
)

The table is already filled with some mappings of MapVal1 to MapVal2. The issue is that not all the mappings should be flagged as Actual. For this reason should be used IsActual column. Currently IsActual is set to NULL for every row. The task is to create the query for updating IsActual column value. UPDATE query should follow next conditions:

1. If MapVal1 is unique and MapVal2 is unique (one-to-one mapping) - then this mapping should be flagged as Actual, so IsActual = 1;
2. If MapVal1 is not unique - then Actual should be the mapping of current MapVal1 to smallest MapVal2, and this MapVal2 must be not mapped to any other MapVal1 that is smaller than current MapVal1;
3. If MapVal2 is not unique - then Actual should be the mapping of current MapVal2 to smallest MapVal1, and this MapVal1 must be not mapped to any other MapVal2 that is smaller than current MapVal2;
4. All rows that are not fulfill any of 1), 2) or 3) conditions - should be flagged as inactual, so IsActual = 0. I believe there is relation between Condition 2) and Condition 3). For every row they both are fulfilled or both are not.

To make it clear, here is an example of result I want to obtain:

Result should be that every MapVal1 is mapped to just one MapVal2 and vice varsa every MapVal2 is mapped to just one MapVal1.

I have created sql-query to resolve my task:

IF OBJECT_ID('tempdb..#TableA') IS NOT NULL
BEGIN  
  DROP TABLE #TableA
END

CREATE TABLE #TableA 
( 
  ID BIGINT IDENTITY (1, 1) PRIMARY KEY,
  MapVal1 BIGINT NOT NULL,
  MapVal2 BIGINT NOT NULL,
  IsActual BIT NULL
)

-- insert input data
INSERT INTO #TableA (MapVal1, MapVal2) 
SELECT 1, 1
UNION ALL SELECT 1, 3
UNION ALL SELECT 1, 4
UNION ALL SELECT 2, 1
UNION ALL SELECT 2, 3
UNION ALL SELECT 2, 4
UNION ALL SELECT 3, 3
UNION ALL SELECT 3, 4
UNION ALL SELECT 4, 3
UNION ALL SELECT 4, 4
UNION ALL SELECT 6, 7
UNION ALL SELECT 7, 8
UNION ALL SELECT 7, 9
UNION ALL SELECT 8, 8
UNION ALL SELECT 8, 9
UNION ALL SELECT 9, 8
UNION ALL SELECT 9, 9


CREATE NONCLUSTERED INDEX IX_Mapping_MapVal1 ON #TableA (MapVal1); 
CREATE NONCLUSTERED INDEX IX_Mapping_MapVal2 ON #TableA (MapVal2); 


-- UPDATE of #TableA is starting here 

-- every one-to-one mapping should be actual
UPDATE m1 SET
  m1.IsActual = 1
FROM #TableA m1
LEFT JOIN #TableA m2
  ON m1.MapVal1 = m2.MapVal1 AND m1.ID <> m2.ID
LEFT JOIN #TableA m3
  ON m1.MapVal2 = m3.MapVal2 AND m1.ID <> m3.ID
WHERE m2.ID IS NULL AND m3.ID IS NULL


-- update for every one-to-many or many-to-many mapping is more complicated
-- would be great to change this part of query to make it witout any LOOP
DECLARE @MapVal1 BIGINT
DECLARE @MapVal2 BIGINT

DECLARE @i BIGINT
DECLARE @iMax BIGINT
DECLARE @LoopCount INT = 0 
SELECT 
  @iMax = MAX (m.ID)
FROM #TableA m

SELECT 
  @i = MIN (m.ID)
FROM #TableA m
WHERE m.IsActual IS NULL

WHILE @i <= @iMax
BEGIN  
  
  SELECT @LoopCount = @LoopCount + 1

  SELECT
    @MapVal1 = m.MapVal1,
    @MapVal2 = m.MapVal2
  FROM #TableA m
  WHERE m.ID = @i

  IF EXISTS 
  (
    SELECT * 
    FROM #TableA m 
    WHERE 
      m.ID < @i 
      AND 
        (m.MapVal1 = @MapVal1 
        OR m.MapVal2 = @MapVal2)
      AND m.IsActual IS NULL     
  ) 
  BEGIN
    UPDATE m SET 
      m.IsActual = 0 
    FROM #TableA m 
    WHERE m.ID = @i
  END

  SELECT @i = MIN (m.ID)
  FROM #TableA m
  WHERE 
    m.ID > @i 
    AND m.IsActual IS NULL 
  
END

UPDATE m SET  
  m.IsActual = 1
FROM #TableA m
WHERE m.IsActual IS NULL


SELECT * FROM #TableA

but as it was expected performance of the query with LOOP is very bad, specially when input table keep millions of rows. I spent a lot of time trying to produce query without LOOP to get reduce execution time of my query but unsuccessfully.

Could anybody advice me how to improve performance of my query. It would be great to get query without LOOP.

Answer 1

Using a loop does not imply you need to update the table one record at a time. It may help if each individual UPDATE statement updates multiple records.

Consider all possible combinations of MapVal1 and MapVal2 as a matrix. Every time you flag a cell as 'actual', you can flag an entire row and an entire column as 'not actual'.

The simplest way to do this, is by following these steps.

1. Of all mappings with IsActual = NULL, take the first one (smallest MapVal1, together with the smallest MapVal2 it is mapped to).
2. Flag this mapping as actual (IsActual = 1).
3. Flag all other mappings with the same MapVal1 as non-actual (IsActual = 0).
4. Flag all other mappings with the same MapVal2 as non-actual (IsActual = 0).
5. Repeat from step 1 until no more records with IsActual = NULL exist.

Here's an implementation:

SELECT 0    -- force @@ROWCOUNT initially 1

WHILE @@ROWCOUNT > 0
    WITH MakeActual AS (
        SELECT TOP 1 MapVal1, MapVal2
        FROM #TableA
        WHERE IsActual IS NULL
        ORDER BY MapVal1, MapVal2
    )
    UPDATE a
    SET IsActual = CASE WHEN a.MapVal1 = m.MapVal1 AND a.MapVal2 = m.MapVal2 THEN 1 ELSE 0 END
    FROM #TableA a
    INNER JOIN MakeActual m ON a.MapVal1 = m.MapVal1 OR a.MapVal2 = m.MapVal2

The number of loop iterations equals the number of 'actual' mappings. The actual performance gain depends a lot on the data. If the majority of mappings is one-to-one (ie hardly any non-actual mappings), then my algorithm will make little difference. Therefore, it may be wise to keep the initial UPDATE statement from your own code sample (the one with the comment "every one-to-one mapping should be actual").

It may also help to play around with the indexes. This one should help to further optimize the clause ORDER BY MapVal1, MapVal2 :

CREATE NONCLUSTERED INDEX IX_MapVals ON #TableA (MapVal1, MapVal2)