How to preset the order of rows in the outer query of a correlated query ?

Question

Good morning,
I have the following simple query:
select empno,
       ename,
       sal,
       sum(case
             when rn = 1 then sal
             else -sal
           end) over (order by sal, empno) as running_diff
   from (
         select empno,
                ename,
                sal,
                row_number() over (order by sal, empno) as rn
           from emp
          where deptno = 10
         );That calculates a running difference and uses "row_number() over (...)" which is an Oracle specific feature to do so. It yields the following result (which we will consider correct):
     EMPNO ENAME             SAL RUNNING_DIFF
      7934 MILLER           1300         1300
      7782 CLARK            2450        -1150
      7839 KING             5000        -6150I wanted to come up with a solution that was not Oracle specific different solution. I tried the following code:
(EDIT: after additional thought, that code is totally different in meaning and will never come close, to the above result. Consider it wrong and ignore this attempt altogether.)
select a.empno,
       a.ename,
       a.sal,
       (select case
                 when a.empno = min(b.empno) then sum(b.sal)
                 else sum(-b.sal)
               end
          from emp b
         where b.empno <= a.empno
           and b.deptno = a.deptno) as running_diff
  from emp a
where a.deptno = 10;but the result is
     EMPNO ENAME             SAL RUNNING_DIFF
      7782 CLARK            2450         2450
      7839 KING             5000        -7450
      7934 MILLER           1300        -8750which is a long way from the original result. I've tried everything I could think of to order the rows before the running difference is calculated but, have been unsuccessful.
Is there a way to change this second query --(without using Oracle specific features)-- without using windowing features that would yield the same result as the first query ?
Rephrase of the above question:
Is there a way, using plain vanilla SQL (that is aggregate functions and set operations such as joins and unions) to create a query that yields the same result as the first one ?
Also, this is not for production code. This is simply an exercise in set manipulation that I'd like to see a solution for.
Thank you for your help,
John.
Edited by: 440bx - 11gR2 on Jul 18, 2010 12:50 AM - correct "ho w" to "How"
Edited by: 440bx - 11gR2 on Jul 18, 2010 1:42 AM - struck out all references to row_number and windowing features being Oracle specific features.
Edited by: 440bx - 11gR2 on Jul 18, 2010 3:51 AM - Noted that my try is woefully wrong and restated the objective to make it clearer.

Answer 1

Hi, John,
One way to get a running total (which is basically what you want) is to do a self-join. Join each row (let's call it the current row, or c) to itself and everything that came before it (let's call this the previous row, or p), and do a regular aggregate SUM, like this:
WITH     got_base_sal     AS
     SELECT       deptno
     ,       2 * MIN (sal)     AS base_sal
     FROM       scott.emp
     GROUP BY  deptno
SELECT       c.deptno
,       c.empno
,       c.ename
,       c.sal
,       b.base_sal - SUM (p.sal)     AS running_diff
FROM       scott.emp     c
JOIN       scott.emp     p     ON     c.deptno     = p.deptno
                    AND     (     c.sal     > p.sal
                         OR     (     c.sal     =  p.sal
                              AND     c.empno     >= p.empno
JOIN       got_base_sal     b     ON     c.deptno     = b.deptno
WHERE       c.deptno     IN (10)
GROUP BY  c.deptno
,       c.empno
,       c.ename
,       c.sal
,       b.base_sal
ORDER BY  c.deptno
,       running_diff     DESC
;Output:
    DEPTNO      EMPNO ENAME             SAL RUNNING_DIFF
        10       7934 MILLER           1300         1300
        10       7782 CLARK            2450        -1150
        10       7839 KING             5000        -6150I said you basically want a runninng total. There are two differences between a running total and your requirements
(1) You want to have a total of the negative of what's in the table. That's trivial: use a minus sign.
(2) You want the first item to count as positive instead of negative. That's not so trivial. The query above counts all sals as negative, but adds an offset so that it appears as if the first item had been counted as positive, not negative.
You didn't say what you want to do in case of a tie (two or more rows having the same sal). The query above uses empno as a tie-breaker, so that all sals are calculated as if they were distinct. This is similar to what analytic functions do when the windowing is based on rows. If you want something similar to windowing by range, that might actually be simpler.
The query above calculates a separate running_diff for each deptno, similar to "PARTITION BY deptno" in analytic functions. You happen to be interested in only one deptno right now, but you can change the main query's WHERE clause, or omit it, and the query will still work. If you don't want this feature (analagoud to not having any PARTITION BY), it's easy to modify the query.
You could also get these results using a recursive WITH clause. That meets the criteria of avoiding analytic functions and Oracle-specific features, but not the one about using only plain, simple SQL features.