Slow performance on Recursive CTE in T-SQL - sql

Question

I have the following code, (based on the community member help):
use [Credible];
WITH DataSource AS
(
SELECT
ROW_NUMBER() OVER(ORDER BY epi.[client_id]) AS [row_id]
,epi.[client_id]
,CONVERT(date, epi.[admission_date]) AS [admission_date]
,CONVERT(date, DATEADD(MONTH, 3, epi.[admission_date])) AS [3Month Date]
,CONVERT(date, ISNULL(epi.[discharge_date], GETDATE())) AS [discharge_date]
FROM
[dbo].[ClientEpisode] epi
WHERE DATEADD(MONTH, 3, [admission_date]) <= ISNULL([discharge_date], GETDATE())
),
RecursiveDataSource AS
(
SELECT
[row_id]
,[client_id]
,[admission_date]
,[3Month Date]
,[discharge_date]
,0 AS [level]
FROM
DataSource
UNION ALL
SELECT
ds.[row_id]
,ds.[client_id]
,ds.[admission_date]
,DATEADD(MONTH, 3, rds.[3Month Date])
,ds.[discharge_date]
,[level] + 1
FROM
RecursiveDataSource rds
INNER JOIN DataSource ds ON
rds.[row_id] = ds.[row_id] AND DATEADD(MONTH, 3, rds.[3Month Date]) < ds.[discharge_date]
)
SELECT *
FROM RecursiveDataSource
ORDER BY [row_id]
,[level]
-- OPTION (MAXRECURSION 32767);
This code works with the speed around 30 sec if there are up to 1 000 records in the table.
But my table is over 14 000 records, and will grow more,
and the code works 10+++ min
Is there way to make its performance in 30 sec or so?
Thank you for help

Answer 1

As here you are looking for performance solution, I will advice to skip using recursive CTE and solve the issue in another way.
For example, a variant of this:
DECLARE #DataSource TABLE
(
[client_id] INT
,[adm_date] DATE
,[disch_date] DATE
);
INSERT INTO #DataSource ([client_id], [adm_date], [disch_date])
VALUES (1002, '3/11/2005 ', '5/2/2005')
,(1002, '8/30/2005 ', '2/16/2007')
,(1002, '3/16/2017 ', NULL);
DROP TABLE IF EXISTS #DataSource;
DROP TABLE IF EXISTS #DataNumers;
CREATE TABLE #DataSource
(
[client_id] INT
,[adm_date] DATE
,[disch_date] DATE
,[diff_in_months] INT
);
CREATE TABLE #DataNumers
(
[number] INT
);
DECLARE #max_diff_in_months INT;
INSERT INTO #DataSource ([client_id], [adm_date], [disch_date], [diff_in_months])
SELECT [client_id]
,[adm_date]
,ISNULL([disch_date], GETUTCDATE()) AS [disch_date]
,CEILING(DATEDIFF(MONTH, [adm_date], ISNULL([disch_date], GETUTCDATE())) * 1.0 / 3.0) - 1
FROM #DataSource
WHERE DATEADD(MONTH, 3, [adm_date]) <= ISNULL([disch_date], GETUTCDATE());
SELECT #max_diff_in_months = MAX([diff_in_months])
FROM #DataSource;
INSERT INTO #DataNumers ([number])
SELECT TOP (#max_diff_in_months) ROW_NUMBER() OVER(ORDER BY (SELECT 1))
FROM [master]..[spt_values];
SELECT DS.[client_id]
,DS.[adm_date]
,DATEADD(MONTH, DN.[number] * 3, [adm_date]) AS [3Month Date]
,DS.[disch_date]
FROM #DataSource DS
CROSS APPLY #DataNumers DN
WHERE DS.[diff_in_months] >= DN.[number];
The algorithm is the following:
select only the rows we needed
calculate the differences in months/3 for each row
generate numbers in helper table
generate the new rows using cross apply
You can use CROSS JOIN if you have more then 2.5k numbers to generate but I remove it in my example. Also, we are using [master]..[spt_values] to generate the total periods. But it's just one way for generating numbers in T-SQL. You can check the link and use another technique if you like.
The idea is to find the periods between each [adm_date] and [disch_date] - and a period is measure by 3 months - that's why we are using DATEDIFF with MONTH and then dividing by 3. We are using ceiling to get the upper value. Of course, you can generate more periods than needed and exclude them with WHERE clause in the final result.