Question

I do struggle with a GROUP BY — again. The basics I can handle, but there it is: How do I get to different columns I named in the group by, without destroying my grouping? Note that group by is only my own idea, there may be others that work better. It must work in Oracle, though.

Here is my example:

create table xxgroups (
  groupid int not null primary key,
  groupname varchar2(10)
);
insert into xxgroups values(100, 'Group 100');
insert into xxgroups values(200, 'Group 200');

drop table xxdata;
create table xxdata (
  num1 int,
  num2 int,
  state_a int,
  state_b int,
  groupid int,
    foreign key (groupid) references xxgroups(groupid)
);
-- "ranks" are 90, 40, null, 70:
insert into xxdata values(10, 10, 1, 4, 100);
insert into xxdata values(10, 10, 0, 4, 200);
insert into xxdata values(11, 11, 0, 3, 100);
insert into xxdata values(20, 22, 5, 7, 200);

The task is to create a result-row for each distinct (num1, num2) and print that groupname with the highest calculated “rank” from state_a and state_b.

Note that the first two rows have the same nums and thus only the higher ranking should be selected — with the groupname being “Group 200”.

I got quite far with the basic group by, I think.

SELECT xd.num1||xd.num2 nummer, max(ranking.goodness)
FROM xxdata xd
    , xxgroups xg
    ,( select state_a, state_b, r as goodness
       from dual
       model return updated rows
       dimension by (0 state_a, 0 state_b) measures (0 r)
       rules (r[1,4]=90, r[3,7]=80,r[5,7]=70, r[4,7]=60, r[0,7]=50, r[0,4]=40)
       order by goodness desc
      ) ranking
WHERE xd.groupid=xg.groupid
  and ranking.state_a (+) = xd.state_a
  and ranking.state_b (+) = xd.state_b
GROUP BY xd.num1||xd.num2
ORDER BY nummer
;

The result is 90% of what I need:

nummer   ranking
----------------
1010     90
1111      
2022     70

100% perfect would be

nummer   groupname
-------------------
1010     Group 100
1111     Group 100
2022     Group 200

• The tricky part is, that I want the groupname in the result. And I can not include it in the select, because then I would have to put it into the group by as well — which I do not want (then I would not select the best ranking entry from over all groups)
• In my solution a use a model table to calculate the “rank”. There are other solution I am sure. The point is, that it is a non-trivial calculation that I do not want to do twice.
• I know from other examples that one could use a second query to get back to the original row to get to the groupname, but I can not see how I could to this here,
  without duplicating my ranking calculation.
• A nice suggestion was to replace the group by with a LIMIT 1/ORDER BY goodness and use this calculating select as a filtering subselect. But a) there is no LIMIT in Oracle, and I doubt a rownum<=1 would do in a subselect and b) I can not wrap my brain around it anyway. Maybe there is a way?

Answer 1

You can use the FIRST aggregation modifier to selectively apply your function over a subset of rows of a group — here a single row (SQLFiddle demo):

SELECT xd.num1||xd.num2 nummer, 
       MAX(xg.groupname) KEEP (DENSE_RANK FIRST 
                               ORDER BY ranking.goodness DESC) grp,
       max(ranking.goodness)
FROM xxdata xd
    , xxgroups xg
    ,( select state_a, state_b, r as goodness
       from dual
       model return updated rows
       dimension by (0 state_a, 0 state_b) measures (0 r)
       rules (r[1,4]=90, r[3,7]=80,r[5,7]=70, r[4,7]=60, r[0,7]=50, r[0,4]=40)
       order by goodness desc
      ) ranking
WHERE xd.groupid=xg.groupid
  and ranking.state_a (+) = xd.state_a
  and ranking.state_b (+) = xd.state_b
GROUP BY xd.num1||xd.num2
ORDER BY nummer;

Your method with analytics works as well but since we already use aggregations here, we may as well use the FIRST modifier to get all columns in one go.