Oracle Analytic functions - resetting a windowing clause

Question

I have the following data set.

create table t1 (
  dept number,
  date1 date
);

Table created.

insert into t1 values (100, \'01-jan-2013\');
insert into t1         


        

Answer 1

The approach is:

1. Mark each row with a 'row number' and a 'changed' flag
2. Calculate the final 'rnk' as the difference between the 'row number', and the maximum previous 'row number' that corresponds to a 'changed' row.

This is similar to Gordon's answer, but written using a CTE which I find easier to read.

with cte as (
    select dept, date1,
        row_number() over (order by date1) as row,
        case when dept = (lag(dept) over (order by date1)) then 0 else 1 end as changed
    from t1
)
select dept, date1, 
    row - max(case when changed = 1 then row else 1 end) over (order by date1) + 1 as rnk
from cte
order by date1

Answer 2

This could have been a case for model clause, but unfortunately it dramatically underperforms on significant amount of rows compared to Gordon's query.

select dept, date1, rank from t1
model 
  dimension by ( row_number() over(order by date1) as rn )
  measures( 1 as rank, dept, date1 ) 
  rules ( 
    rank[1] = 1,
    rank[rn > 1] = 
    case dept[cv()] 
      when dept[cv()-1] then rank[cv()-1] + 1 
      else 1
     end
  )

http://www.sqlfiddle.com/#!4/fc339/132

Answer 3

This is a little complicated. Instead of using rank() or the like, use lag() to see when something changes. Then do a cumulative sum of the flag.

select dept, date1,
       CASE WHEN StartFlag = 0 THEN 1
            ELSE 1+StartFlag+NVL(lag(StartFlag) over (order by date1),0)
       END as rnk
from (select t1.*,
             (case when dept = lag(dept) over (order by date1)
                   then 1
                   else 0
              end) as StartFlag
      from t1
     ) t1
order by date1;

Here is the SQLFiddle.

EDIT:

This is Gordon editing my own answer. Oops. The original query was 90% of the way there. It identified the groups where the numbers should increase, but did not assign the numbers within the groups. I would do this with another level of row_number() as in:

select dept, date1,
       row_number() over (partition by dept, grp order by date1) as rnk
from (select dept, date1, startflag,
             sum(StartFlag) over (partition by dept order by date1) as grp
      from (select t1.*,
                   (case when dept = lag(dept) over (order by date1)
                         then 0
                         else 1
                    end) as StartFlag
            from t1
           ) t1
     ) t1
order by date1;

So, the overall idea is the following. First use lag() to determine where a group begins (that is, where there is a department change from one date to the next). Then, assign a "group id" to these, by doing a cumulative sum. These are the records that are to be enumerated. The final step is to enumerate them using row_number().