Left join with nearest value without duplicates - sql

Question

I want to achieve in MS SQL something like below, using 2 tables and through join instead of iteration.
From table A, I want each row to identify from table B which in the list is their nearest value, and when value has been selected, that value cannot re-used. Please help if you've done something like this before. Thank you in advance! #SOreadyToAsk

Answer 1

Below is a set-based solution using CTEs and windowing functions.
The ranked_matches CTE assigns a closest match rank for each row in TableA along with a closest match rank for each row in TableB, using the index value as a tie breaker.
The best_matches CTE returns rows from ranked_matches that have the best rank (rank value 1) for both rankings.
Finally, the outer query uses a LEFT JOIN from TableA to the to the best_matches CTE to include the TableA rows that were not assigned a best match due to the closes match being already assigned.
Note that this does not return a match for the index 3 TableA row indicated in your sample results. The closes match for this row is TableB index 3, a difference of 83. However, that TableB row is a closer match to the TableA index 2 row, a difference of 14 so it was already assigned. Please clarify you question if this isn't what you want. I think this technique can be tweaked accordingly.
CREATE TABLE dbo.TableA(
[index] int NOT NULL
CONSTRAINT PK_TableA PRIMARY KEY
, value int
);
CREATE TABLE dbo.TableB(
[index] int NOT NULL
CONSTRAINT PK_TableB PRIMARY KEY
, value int
);
INSERT INTO dbo.TableA
( [index], value )
VALUES ( 1, 123 ),
( 2, 245 ),
( 3, 342 ),
( 4, 456 ),
( 5, 608 );
INSERT INTO dbo.TableB
( [index], value )
VALUES ( 1, 152 ),
( 2, 159 ),
( 3, 259 );
WITH
ranked_matches AS (
SELECT
a.[index] AS a_index
, a.value AS a_value
, b.[index] b_index
, b.value AS b_value
, RANK() OVER(PARTITION BY a.[index] ORDER BY ABS(a.Value - b.value), b.[index]) AS a_match_rank
, RANK() OVER(PARTITION BY b.[index] ORDER BY ABS(a.Value - b.value), a.[index]) AS b_match_rank
FROM dbo.TableA AS a
CROSS JOIN dbo.TableB AS b
)
, best_matches AS (
SELECT
a_index
, a_value
, b_index
, b_value
FROM ranked_matches
WHERE
a_match_rank = 1
AND b_match_rank= 1
)
SELECT
TableA.[index] AS a_index
, TableA.value AS a_value
, best_matches.b_index
, best_matches.b_value
FROM dbo.TableA
LEFT JOIN best_matches ON
best_matches.a_index = TableA.[index]
ORDER BY
TableA.[index];
EDIT:
Although this method uses CTEs, recursion is not used and is therefore not limited to 32K recursions. There may be room for improvement here from a performance perspective, though.

Answer 2

I don't think it is possible without a cursor.
Even if it is possible to do it without a cursor, it would definitely require self-joins, maybe more than once. As a result performance is likely to be poor, likely worse than straight-forward cursor. And it is likely that it would be hard to understand the logic and later maintain this code. Sometimes cursors are useful.
The main difficulty is this part of the question:
when value has been selected, that value cannot re-used.
There was a similar question just few days ago.
The logic is straight-forward. Cursor loops through all rows of table A and with each iteration adds one row to the temporary destination table. To determine the value to add I use EXCEPT operator that takes all values from the table B and removes from them all values that have been used before. My solution assumes that there are no duplicates in value in table B. EXCEPT operator removes duplicates. If values in table B are not unique, then temporary table would hold unique indexB instead of valueB, but main logic remains the same.
Here is SQL Fiddle.
Sample data
DECLARE #TA TABLE (idx int, value int);
INSERT INTO #TA (idx, value) VALUES
(1, 123),
(2, 245),
(3, 342),
(4, 456),
(5, 608);
DECLARE #TB TABLE (idx int, value int);
INSERT INTO #TB (idx, value) VALUES
(1, 152),
(2, 159),
(3, 259);
Main query inserts result into temporary table #TDst. It is possible to write that INSERT without using explicit variable #CurrValueB, but it looks a bit cleaner with variable.
DECLARE #TDst TABLE (idx int, valueA int, valueB int);
DECLARE #CurrIdx int;
DECLARE #CurrValueA int;
DECLARE #CurrValueB int;
DECLARE #iFS int;
DECLARE #VarCursor CURSOR;
SET #VarCursor = CURSOR FAST_FORWARD
FOR
SELECT idx, value
FROM #TA
ORDER BY idx;
OPEN #VarCursor;
FETCH NEXT FROM #VarCursor INTO #CurrIdx, #CurrValueA;
SET #iFS = ##FETCH_STATUS;
WHILE #iFS = 0
BEGIN
SET #CurrValueB =
(
SELECT TOP(1) Diff.valueB
FROM
(
SELECT B.value AS valueB
FROM #TB AS B
EXCEPT -- remove values that have been selected before
SELECT Dst.valueB
FROM #TDst AS Dst
) AS Diff
ORDER BY ABS(Diff.valueB - #CurrValueA)
);
INSERT INTO #TDst (idx, valueA, valueB)
VALUES (#CurrIdx, #CurrValueA, #CurrValueB);
FETCH NEXT FROM #VarCursor INTO #CurrIdx, #CurrValueA;
SET #iFS = ##FETCH_STATUS;
END;
CLOSE #VarCursor;
DEALLOCATE #VarCursor;
SELECT * FROM #TDst ORDER BY idx;
Result
idx valueA valueB
1 123 152
2 245 259
3 342 159
4 456 NULL
5 608 NULL
It would help to have the following indexes:
TableA - (idx) include (value), because we SELECT idx, value ORDER BY idx;
TableB - (value) unique, Temp destination table - (valueB) unique filtered NOT NULL, to help EXCEPT. So, it may be better to have a temporary #table for result (or permanent table) instead of table variable, because table variables can't have indexes.
Another possible method would be to delete a row from table B (from original or from a copy) as its value is inserted into result. In this method we can avoid performing EXCEPT again and again and it could be faster overall, especially if it is OK to leave table B empty in the end. Still, I don't see how to avoid cursor and processing individual rows in sequence.
SQL Fiddle
DECLARE #TDst TABLE (idx int, valueA int, valueB int);
DECLARE #CurrIdx int;
DECLARE #CurrValueA int;
DECLARE #iFS int;
DECLARE #VarCursor CURSOR;
SET #VarCursor = CURSOR FAST_FORWARD
FOR
SELECT idx, value
FROM #TA
ORDER BY idx;
OPEN #VarCursor;
FETCH NEXT FROM #VarCursor INTO #CurrIdx, #CurrValueA;
SET #iFS = ##FETCH_STATUS;
WHILE #iFS = 0
BEGIN
WITH
CTE
AS
(
SELECT TOP(1) B.idx, B.value
FROM #TB AS B
ORDER BY ABS(B.value - #CurrValueA)
)
DELETE FROM CTE
OUTPUT #CurrIdx, #CurrValueA, deleted.value INTO #TDst;
FETCH NEXT FROM #VarCursor INTO #CurrIdx, #CurrValueA;
SET #iFS = ##FETCH_STATUS;
END;
CLOSE #VarCursor;
DEALLOCATE #VarCursor;
SELECT
A.idx
,A.value AS valueA
,Dst.valueB
FROM
#TA AS A
LEFT JOIN #TDst AS Dst ON Dst.idx = A.idx
ORDER BY idx;

Answer 3

I highly believe THIS IS NOT A GOOD PRACTICE because I am bypassing the policy SQL made for itself that functions with side-effects (INSERT,UPDATE,DELETE) is a NO, but due to the fact that I want solve this without resulting to iteration options, I came up with this and gave me better view of things now.
create table tablea
(
num INT,
val MONEY
)
create table tableb
(
num INT,
val MONEY
)
I created a hard-table temp which I shall drop from time-to-time.
if((select 1 from sys.tables where name = 'temp_tableb') is not null) begin drop table temp_tableb end
select * into temp_tableb from tableb
I created a function that executes xp_cmdshell (this is where the side-effect bypassing happens)
CREATE FUNCTION [dbo].[GetNearestMatch]
(
#ParamValue MONEY
)
RETURNS MONEY
AS
BEGIN
DECLARE #ReturnNum MONEY
, #ID INT
SELECT TOP 1
#ID = num
, #ReturnNum = val
FROM temp_tableb ORDER BY ABS(val - #ParamValue)
DECLARE #SQL varchar(500)
SELECT #SQL = 'osql -S' + ##servername + ' -E -q "delete from test..temp_tableb where num = ' + CONVERT(NVARCHAR(150),#ID) + ' "'
EXEC master..xp_cmdshell #SQL
RETURN #ReturnNum
END
and my usage in my query simply looks like this.
-- initialize temp
if((select 1 from sys.tables where name = 'temp_tableb') is not null) begin drop table temp_tableb end
select * into temp_tableb from tableb
-- query nearest match
select
*
, dbo.GetNearestMatch(a.val) AS [NearestValue]
from tablea a
and gave me this..