Question

I have a huge difference of time execution between a 1-minute query and the same one in a table-valued function.

But the most weired thing is that running the UDF with another (valid) company_id argument
gives me a result in ~40 seconds and as soon as I change this company_id for 12 (valid again), it never stops. The execution plans of these two queries are absolutely not the same and of course, the long one is the most complicated. BUT the execution plan between the batch version and the UDF version are the same AND the batch version is fast…!

If I do the following query “by hand”, the execution time is 1min36s with 306 rows:

 SELECT
  dbo.date_only(Call.date) AS date,
  count(DISTINCT customer_id) AS new_customers
 FROM
  Call
 LEFT OUTER JOIN
  dbo.company_new_customers(12, 2009, 2009) new_customers
   ON dbo.date_only(new_customers.date) = dbo.date_only(Call.date)
 WHERE
  company_id = 12
  AND year(Call.date) >= 2009
  AND year(Call.date) <= 2009
 GROUP BY
  dbo.date_only(Call.date)

I stored this exactly same query in a function and ran it like that :

SELECT * FROM company_new_customers_count(12, 2009, 2009)

13 minutes for now that it is running… And I am sure that it will never give me any result.

Yesterday, I had the exact same infinite-loop-like behaviour during more than 4h (so I stopped it).

Here is the definition of the function:

SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
CREATE FUNCTION company_new_customers_count 
( 
 @company_id int, 
 @start_year int,
 @end_year int
)
RETURNS TABLE 
AS
RETURN 
(
 SELECT
  dbo.date_only(Call.date) AS date,
  count(DISTINCT customer_id) AS new_customers
 FROM
  Call
 LEFT OUTER JOIN
  dbo.company_new_customers(@company_id, @start_year, @end_year) new_customers
   ON dbo.date_only(new_customers.date) = dbo.date_only(Call.date)
 WHERE
  company_id = @company_id
  AND year(Call.date) >= @start_year
  AND year(Call.date) <= @end_year
 GROUP BY
  dbo.date_only(Call.date)
)
GO

I would be very happy to understand what is going on.

Thanks

Additional:

Definition of company_new_customers:

SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
-- =============================================
-- Description: Create the list of new customers of @company_id
--          in the given period.
-- =============================================
CREATE FUNCTION company_new_customers 
(   
    @company_id int, 
    @start_year int,
    @end_year   int
)
RETURNS TABLE 
AS
RETURN 
(
    SELECT
        customer_id,
        date
    FROM
    (   -- select apparition dates of cutomers before @end_year
        SELECT
            min(date)       AS date,
            customer_id
        FROM
            Call
        JOIN
            Call_Customer ON Call_Customer.call_id = Call.call_id
        WHERE
            company_id = @company_id
            AND year(date) <= @end_year
        GROUP BY
            customer_id
    ) new_customers
    WHERE
        year(date) >= @start_year -- select apparition dates of cutomers after @start_year
)
GO

Definition of date_only:

SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
-- =============================================
-- Author:      Julio Guerra
-- Create date: 14/10/2010
-- Description: Return only the date part of a datetime value
-- Example:         date_only('2010-10-25 13:00:12') returns 2010-10-25
-- =============================================
CREATE FUNCTION date_only
(
    @datetime datetime
)
RETURNS datetime
AS
BEGIN
    RETURN dateadd(dd, 0, datediff(dd, 0, @datetime))
END
GO

Execution Plan of SELECT * FROM company_new_customers_count(8, 2009, 2009)

Execution Plan of SELECT * FROM company_new_customers_count(12, 2009, 2009)

Answer 1

From these query plans it looks like you could benefit from an index like this (if I inferred your DB schema right):

CREATE INDEX IX_call_company_date ON call (company_id, date)

In general this seems to be a standard query optimization problem and the table-valued functions aren’t making a difference here actually.