SQL Recursive CTE replace statement too slow - sql

Question

I have a recursive CTE that replaces multiple values from an expression, but it is too slow when there are many expressions.
CREATE TABLE #table1(IdExpresion INT, expresion VARCHAR(MAX))
CREATE TABLE #table2(IdExpresion INT, searchExpresion VARCHAR(50), replacementExpresion VARCHAR(50))
INSERT INTO #table1(IdExpresion, expresion)
VALUES(1, 'Mary had a little lamb'),
(2, 'The new student, student_name has the following grades Math - math_grade, Science - Science_grade')
INSERT INTO #table2(IdExpresion, searchExpresion, replacementExpresion)
VALUES(1, 'lamb','dog'),
(2, 'student_name','Joe Smith'),
(2, 'math_grade','A'),
(2, 'Science_grade','B+')
;WITH cte(IdExpresion, expresion, lvl) AS
(
SELECT t1.IdExpresion, t1.expresion, 1
FROM #table1 t1
UNION ALL
SELECT cte.IdExpresion, REPLACE(cte.expresion, t2.searchExpresion, t2.replacementExpresion), cte.lvl + 1
FROM cte
INNER JOIN #table2 t2
ON cte.IdExpresion = t2.IdExpresion
AND CHARINDEX(t2.searchExpresion, cte.expresion) > 0
)
SELECT DISTINCT c2.expresion
FROM (SELECT IdExpresion, MAX(lvl) AS lvl
FROM cte
GROUP BY IdExpresion) c1
INNER JOIN cte c2
ON c1.IdExpresion = c2.IdExpresion
AND c1.lvl = c2.lvl
OPTION (MAXRECURSION 0);
Anyone have any advice? I am using SQL Server by the way

Answer 1

Not sure if any more performant, but here is a brute force approach just for fun.
Already +1 LukStorm's answer, I suspect that is the way to go.
Example
Declare #S varchar(max) = (Select IdExpresion,expresion = replace(' '+expresion,' ',concat(' ',IdExpresion,'|||')) From #Table1 For XML Raw )
Select #S = replace(#S,concat(IdExpresion,'|||',searchExpresion),replacementExpresion) From #table2
Select IdExpresion = B.i.value('#IdExpresion', 'int')
,expresion = ltrim(replace(B.i.value('#expresion', 'varchar(max)'),B.i.value('#IdExpresion', 'varchar(25)')+'|||',''))
From (Select x = Cast(#S as xml).query('.')) as A
Cross Apply x.nodes('row') AS B(i)
Returns
IdExpresion expresion
1 Mary had a little dog
2 The new student, Joe Smith has the following grades Math - A, Science - B+

Answer 2

You could add another CTE to it that gets a row_number for each replacement, partitioned by the IdExpresion.
Then in the recursive CTE, instead of counting up, count down till there's no match with the replacement row_number.
The last entry in the CTE, that had all replacements, will have Lvl 0 then.
;WITH SEARCH AS (
SELECT
IdExpresion,
row_number() over (partition by IdExpresion order by searchExpresion) as rn,
searchExpresion, replacementExpresion
FROM #table2
), CTE(IdExpresion, expresion, lvl) AS
(
SELECT t1.IdExpresion, t1.expresion, count(*)
FROM #table1 t1
JOIN #table2 t2 ON t2.IdExpresion = t1.IdExpresion
GROUP BY t1.IdExpresion, t1.expresion
UNION ALL
SELECT c.IdExpresion, REPLACE(c.expresion, s.searchExpresion, s.replacementExpresion), c.lvl - 1
FROM CTE c
JOIN SEARCH s
ON s.IdExpresion = c.IdExpresion AND s.rn = c.lvl
)
SELECT IdExpresion, expresion
FROM CTE
WHERE lvl = 0
OPTION (MAXRECURSION 0);
This way, each REPLACE is only done once per IdExpresion.
And that without having to use CHARINDEX.
You could also replace that SEARCH cte with a temporary table.
One that has the records from #table2 with that row_number.
This has the benefit that with a table you can add a compound index.
On a large table it should speed up the recursive join to the replacements.
Test on rextester here
CREATE TABLE #tmpSearch (
IdExpresion INT,
rn INT,
searchExpresion VARCHAR(50),
replacementExpresion VARCHAR(50),
primary key (IdExpresion, rn));
insert into #tmpSearch (IdExpresion, rn, searchExpresion, replacementExpresion)
select
IdExpresion,
row_number() over (partition by IdExpresion order by searchExpresion) as rn,
searchExpresion,
replacementExpresion
from #table2
order by IdExpresion, searchExpresion;
;WITH CTE(IdExpresion, expresion, lvl) AS
(
SELECT t1.IdExpresion, t1.expresion, max(s.rn)
FROM #table1 t1
JOIN #tmpSearch s ON s.IdExpresion = t1.IdExpresion
GROUP BY t1.IdExpresion, t1.expresion
UNION ALL
SELECT c.IdExpresion, REPLACE(c.expresion, s.searchExpresion, s.replacementExpresion), c.lvl - 1
FROM CTE c
JOIN #tmpSearch s
ON s.IdExpresion = c.IdExpresion AND s.rn = c.lvl
)
SELECT IdExpresion, expresion
FROM CTE
WHERE lvl = 0
OPTION (MAXRECURSION 0);

Answer 3

Good day,
Here is another solution. Please check if this fit your needs. This solution does not use any loop but simple dynamic query.
DECLARE #SQLString nvarchar(MAX);
-- do not make mistake, this is simple CTE and not a recursive CTE (no Loop)
;With MyCTE as (
select R
From table1 t1
CROSS APPLY (
SELECT R = 'SELECT ' + CONVERT (NVARCHAR(MAX),t1.IdExpresion) + ' as IdExpresion,' + STRING_AGG ('REPLACE','(') + '(' + 't1.expresion,''' + STRING_AGG(t2.searchExpresion + ''',''' + t2.replacementExpresion , '''),''') + ''') as expresion FROM table1 t1 where t1.IdExpresion = ' + CONVERT (NVARCHAR(MAX),t1.IdExpresion)
from table2 t2
where t2.IdExpresion = t1.IdExpresion
) C
)
SELECT #SQLString = STRING_AGG(R,'
UNION ALL
')
FROM MyCTE
--PRINT #SQLString
EXECUTE sp_executesql #SQLString
GO
Note! I recommend to execute some tests to confirm that this solves all cases
Note! I am using the function STRING_AGG which was added to SQL Server 2017. In older version you can get the exact same solution using FOR XML statement.
Since we don't have the real DDL+DML we cannot really discuss about performance, but the difference in the execution plans of the solutions is 10% to 90% (In general, You should check IO and Time statistics in production in addition, before choosing your solution)
So... here is the Execution Plans Image (above query is my dynamic SQL solution and bellow is LukStorms solution using recursive CTE = Loop)