Optimal way to concatenate/aggregate strings

Are methods using FOR XML PATH like below really that slow? Itzik Ben-Gan writes that this method has good performance in his T-SQL Querying book (Mr. Ben-Gan is a trustworthy source, in my view).

create table #t (id int, name varchar(20))

insert into #t
values (1, 'Matt'), (1, 'Rocks'), (2, 'Stylus')

select  id
        ,Names = stuff((select ', ' + name as [text()]
        from #t xt
        where xt.id = t.id
        for xml path('')), 1, 2, '')
from #t t
group by id

STRING_AGG() in SQL Server 2017, Azure SQL, and PostgreSQL: https://www.postgresql.org/docs/current/static/functions-aggregate.html
https://docs.microsoft.com/en-us/sql/t-sql/functions/string-agg-transact-sql

GROUP_CONCAT() in MySQL
http://dev.mysql.com/doc/refman/5.7/en/group-by-functions.html#function_group-concat

(Thanks to @Brianjorden and @milanio for Azure update)

Example Code:

select Id
, STRING_AGG(Name, ', ') Names 
from Demo
group by Id

SQL Fiddle: http://sqlfiddle.com/#!18/89251/1


Although @serge answer is correct but i compared time consumption of his way against xmlpath and i found the xmlpath is so faster. I'll write the compare code and you can check it by yourself. This is @serge way:

DECLARE @startTime datetime2;
DECLARE @endTime datetime2;
DECLARE @counter INT;
SET @counter = 1;

set nocount on;

declare @YourTable table (ID int, Name nvarchar(50))

WHILE @counter < 1000
BEGIN
    insert into @YourTable VALUES (ROUND(@counter/10,0), CONVERT(NVARCHAR(50), @counter) + 'CC')
    SET @counter = @counter + 1;
END

SET @startTime = GETDATE()

;WITH Partitioned AS
(
    SELECT 
        ID,
        Name,
        ROW_NUMBER() OVER (PARTITION BY ID ORDER BY Name) AS NameNumber,
        COUNT(*) OVER (PARTITION BY ID) AS NameCount
    FROM @YourTable
),
Concatenated AS
(
    SELECT ID, CAST(Name AS nvarchar) AS FullName, Name, NameNumber, NameCount FROM Partitioned WHERE NameNumber = 1

    UNION ALL

    SELECT 
        P.ID, CAST(C.FullName + ', ' + P.Name AS nvarchar), P.Name, P.NameNumber, P.NameCount
    FROM Partitioned AS P
        INNER JOIN Concatenated AS C ON P.ID = C.ID AND P.NameNumber = C.NameNumber + 1
)
SELECT 
    ID,
    FullName
FROM Concatenated
WHERE NameNumber = NameCount

SET @endTime = GETDATE();

SELECT DATEDIFF(millisecond,@startTime, @endTime)
--Take about 54 milliseconds

And this is xmlpath way:

DECLARE @startTime datetime2;
DECLARE @endTime datetime2;
DECLARE @counter INT;
SET @counter = 1;

set nocount on;

declare @YourTable table (RowID int, HeaderValue int, ChildValue varchar(5))

WHILE @counter < 1000
BEGIN
    insert into @YourTable VALUES (@counter, ROUND(@counter/10,0), CONVERT(NVARCHAR(50), @counter) + 'CC')
    SET @counter = @counter + 1;
END

SET @startTime = GETDATE();

set nocount off
SELECT
    t1.HeaderValue
        ,STUFF(
                   (SELECT
                        ', ' + t2.ChildValue
                        FROM @YourTable t2
                        WHERE t1.HeaderValue=t2.HeaderValue
                        ORDER BY t2.ChildValue
                        FOR XML PATH(''), TYPE
                   ).value('.','varchar(max)')
                   ,1,2, ''
              ) AS ChildValues
    FROM @YourTable t1
    GROUP BY t1.HeaderValue

SET @endTime = GETDATE();

SELECT DATEDIFF(millisecond,@startTime, @endTime)
--Take about 4 milliseconds

SOLUTION

The definition of optimal can vary, but here's how to concatenate strings from different rows using regular Transact SQL, which should work fine in Azure.

;WITH Partitioned AS
(
    SELECT 
        ID,
        Name,
        ROW_NUMBER() OVER (PARTITION BY ID ORDER BY Name) AS NameNumber,
        COUNT(*) OVER (PARTITION BY ID) AS NameCount
    FROM dbo.SourceTable
),
Concatenated AS
(
    SELECT 
        ID, 
        CAST(Name AS nvarchar) AS FullName, 
        Name, 
        NameNumber, 
        NameCount 
    FROM Partitioned 
    WHERE NameNumber = 1

    UNION ALL

    SELECT 
        P.ID, 
        CAST(C.FullName + ', ' + P.Name AS nvarchar), 
        P.Name, 
        P.NameNumber, 
        P.NameCount
    FROM Partitioned AS P
        INNER JOIN Concatenated AS C 
                ON P.ID = C.ID 
                AND P.NameNumber = C.NameNumber + 1
)
SELECT 
    ID,
    FullName
FROM Concatenated
WHERE NameNumber = NameCount

EXPLANATION

The approach boils down to three steps:

  1. Number the rows using OVER and PARTITION grouping and ordering them as needed for the concatenation. The result is Partitioned CTE. We keep counts of rows in each partition to filter the results later.

  2. Using recursive CTE (Concatenated) iterate through the row numbers (NameNumber column) adding Name values to FullName column.

  3. Filter out all results but the ones with the highest NameNumber.

Please keep in mind that in order to make this query predictable one has to define both grouping (for example, in your scenario rows with the same ID are concatenated) and sorting (I assumed that you simply sort the string alphabetically before concatenation).

I've quickly tested the solution on SQL Server 2012 with the following data:

INSERT dbo.SourceTable (ID, Name)
VALUES 
(1, 'Matt'),
(1, 'Rocks'),
(2, 'Stylus'),
(3, 'Foo'),
(3, 'Bar'),
(3, 'Baz')

The query result:

ID          FullName
----------- ------------------------------
2           Stylus
3           Bar, Baz, Foo
1           Matt, Rocks