Accumulative Distinct Counts

Question

I’ve been at this for two days now. Could someone lead me down the right path? Given the following data set:
create table data_owner.test_data
item_number varchar2(10 byte),
store_number varchar2(10 byte),
calendar_year varchar2(10 byte),
calendar_week varchar2(10 byte),
units_sold integer
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '31', '2010', '51', 4)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '16', '2010', '51', 2)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '31', '2010', '52', 3)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '27', '2010', '52', 1)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '16', '2011', '1', 3)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '27', '2011', '2', 5)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('1111', '20', '2011', '2', 4)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('2222', '27', '2010', '51', 3)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('2222', '16', '2010', '52', 2)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('2222', '20', '2010', '52', 1)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('2222', '16', '2011', '1', 3)
insert into test_data(item_number, store_number, calendar_year, calendar_week, units_sold)
values ('2222', '31', '2011', '2', 3)
select * from test_data
item_number     store_number     calendar_year     calendar_week     units_sold
1111     31     2010     51     4
1111     16     2010     51     2
1111     31     2010     52     3
1111     27     2010     52     1
1111     16     2011     1     3
1111     27     2011     2     5
1111     20     2011     2     4
2222     27     2010     51     3
2222     16     2010     52     2
2222     20     2010     52     1
2222     16     2011     1     3
2222     31     2011     2     3
My desired result is a sum of units sold and an accumulative distinct count of store numbers grouped by item, year, and week. i.e.:
item_number     calendar_year     calendar_week     store_count     sum(units_sold)
1111     2010     51     2     6
1111     2010     52     3     4
1111     2011     1     3     3
1111     2011     2     4     9
2222     2010     51     1     3
2222     2010     52     3     3
2222     2011     1     3     3
2222     2011     2     4     3
I can’t seem to get the store count right. I’ve been trying various methods of the count(distinct store_number) over (…) analytic function, but nothing works. Thanks.

Answer 1

Hi,
Interesting problem!
When using analytic functions, you can't use both DISTINCT and ORDER BY. Too bad; that sure would be convenient.
The most general solution is to use aggregate functions instead of analytic functions, and do a self-join to pair every row ("table" l, for "later" in the query below) with every earleir row ("table" e below) for the same item:
SELECT       l.item_number
,       l.calendar_year
,       l.calendar_week
,       COUNT (DISTINCT e.store_number)     AS store_count
,       SUM (l.units_sold)
     / COUNT (DISTINCT e.ROWID)          AS total_units_sold
FROM       test_data   e
JOIN       test_data   l      ON     e.item_number     = l.item_number
AND                               e.calendar_year || LPAD (e.calendar_week, 2)
                                <= l.calendar_year || LPAD (l.calendar_week, 2)
GROUP BY  l.item_number
,       l.calendar_year
,       l.calendar_week
ORDER BY  l.item_number
,       l.calendar_year
,       l.calendar_week
;You might think about storing a DATE (say, the date when the week begins) instead of year and week. It would simplify this query, and probably lots of other ones, too. I realize that might complicate some other queries, but I think you'll fiond a net gain.
Thanks for posting the CREATE TABLE and INSERT statements; that helps a lot!
Edited by: Frank Kulash on Nov 18, 2011 12:48 PM
Here's an analytic solution. As you can see, it requires more code, and more complicated code, but it might perform better:
WITH     got_r_num   AS
     SELECT     item_number
     ,     calendar_year
     ,     calendar_week
     ,     units_sold
     ,     ROW_NUMBER () OVER ( PARTITION BY  item_number
                               ,                    store_number
                         ORDER BY        calendar_year
                         ,                calendar_week
                       )      AS r_num
     FROM    test_data
SELECT DISTINCT
     item_number
,     calendar_year
,     calendar_week
,     COUNT ( CASE
                    WHEN  r_num = 1
              THEN  1
                END
           )             OVER ( PARTITION BY  item_number
                                  ORDER BY      calendar_year
                      ,          calendar_week
                             )                    AS store_count
,       SUM (units_sold) OVER ( PARTITION BY  item_number
                                ,             calendar_year
                      ,             calendar_week
                         )                         AS  total_units_sold
FROM       got_r_num
ORDER BY  item_number
,            calendar_year
,       calendar_week
;This approah will not work in all windowing situations. It's okay fo this job, but not if you wanted,for example, a count of distinct stores from the last 6 weeks, and the report covers more than 6 weeks.