SQL group data (find data family) - sql

Question

Please help me, I need to find out a SQL solution for grouping data using SQL Server database.
I'm pretty sure that it could be done in one SQL request but I can't see the trick.
Let' see the problem :
I have a two columns table (please see below an example). I just want to add a new column containing a number or a string which indicates the group
BEFORE :
Col1 | Col2
-----+-----
A | B
B | C
D | E
F | G
G | H
I | I
J | U
AFTER TRANSFORMATION :
Col1 | Col2 | Group
-----+------+------
A | B | 1
B | C | 1
D | E | 2
F | G | 3
G | H | 3
I | I | 4
J | U | 5
In other words: A, B, C are in the same group; D and E too; F, G, H in group 3 ....

Answer 1

Do you have any lookup table to get this group mapping?
Or if you just have a logic defined to decide a group, i would recommend to add a UDF which will return group for supplied values.
SELECT Col1,Col2,GetGroupID(Col1,Col2) AS Group
FROM Table
Your UDF will be something like following
CREATE FUNCTION GetGroupID
(
-- Add the parameters for the function here
#Col1 varchar(10),
#Col2 varchar(10)
)
RETURNS int
AS
BEGIN
DECLARE #groupID int
IF (#Col1="A" AND #Co2 = "B") OR (#Col1="B" AND #Co2 = "C")
BEGIN
SET #groupID = 1
END
IF #Col1="D" AND #Co2 = "E"
BEGIN
SET #groupID = 2
END
-- You can write saveral conditions in the same manner.
return #groupID
END
However, in case you have this mapping defined somewhere in another table, let us know the structure of the table and we can then update the query to join with that table instead of using UDF.
Considering the performance of the query, if the amount of data is huge in your table , it is recommended to have these mappings to one fix table and Join that table in query. Using UDF may harm performance if data amount is huge.

Answer 2

There is absolutely no need for a UDF here. Regardless of whether you are looking to update the table with a new column or simply pull out the data with the grouping applied, you will be best off using a set based solution, ie: create and join to a table.
I am assuming here that you don't have messy data, such as a row with Col1 = 'A' and Col2 = 'F'.
If you are able to add new tables permanently you can use the following to create your lookup table:
create table Col1Groups(Col1 nvarchar(10), GroupNum int);
insert into Col1Groups(Col1,GroupNum) values ('A',1),('B',1),('C',1),('D',2),('E',2),('F',3),('G',3),('H',3);
and then join to it:
select t.Col1
,t.Col2
,g.GroupNum
from Table t
inner join Col1Groups g
on t.Col1 = g.Col1
If you can't, you can just create a derived table via a CTE:
with Col1Groups as
(
select Col1
,GroupNum
from (values('A',1),('B',1),('C',1),('D',2),('E',2),('F',3),('G',3),('H',3)) as x(Col1,GroupNum)
)
select t.Col1
,t.Col2
,g.GroupNum
from Table t
inner join Col1Groups g
on t.Col1 = g.Col1

Answer 3

You get the first rows per group with
select col1, col2 from mytable where col1 not in (select col2 from mytable) or col1 = col2;
We can give these rows numbers with
rank() over (order by col1) as grp
Now we must iterate through the rows to find the ones belonging to those first ones, then those belonging to these, etc. A recursive query.
with cte(col1, col2, grp) as
(
select col1, col2, rank() over (order by col1) as grp
from mytable where col1 not in (select col2 from mytable) or col1 = col2
union all
select mytable.col1, mytable.col2, cte.grp
from cte
join mytable on mytable.col1 = cte.col2
where mytable.col1 <> mytable.col2
)
select * from cte
order by grp, col1;

Answer 4

Additional answer for a more flexible approach
Originally you asked for chains A|B -> B|C, F|G -> G|H etc., but in your comment to my other answer you introduced forks like A|B -> B|C, B|D and I've adjusted my answer.
If you want to go one step further and introduce net-like relations such as A|B -> B|C, D|C, we can no longer follow chains forward only (in the example D belongs to the A group, because though A doesn't lead to D directly, it leads to C and D also leads to C. Here is a way to solve this:
Get all letters from the table (no matter whether in col1 or col2). Then for each of them find related letters (again no matter whether in col1 or col2). And for these again find related letters and so on. That will give you complete groups. But duplicates (as D is in the A group, A is in the D group also), which you can get rid of by simply taking the smallest (or greatest) group key per letter. Then join the Groups to the table.
The query:
with cte(col, grp) as
(
select col, rownum as grp from
(select col1 as col from mytable union select col2 from mytable)
union all
select case when mytable.col1 = cte.col then mytable.col2 else mytable.col1 end, cte.grp
from cte
join mytable on cte.col in (mytable.col1, mytable.col2)
where mytable.col1 <> mytable.col2
)
cycle col set is_cycle to 'y' default 'n'
select mytable.col1, mytable.col2, x.grp
from mytable
join (select col, min(grp) as grp from cte group by col) x on x.col = mytable.col1
order by grp, col;