Oracle values change when row_number() added

Question

I have an Oracle query that uses DUAL to produce a list of dates in a subquery, and a case when to identify business days:

    SELECT DATES
        ,case when to_char(DATES, 'd') in (1,7)
         then 0
        else 1 end as business_day
    FROM (
        SELECT to_date('1/1/2020','MM/DD/YYYY') + (LEVEL -1) AS DATES
        FROM DUAL connect by level <=(to_date('1/1/2021','MM/DD/YYYY') - to_date('1/1/2020','MM/DD/YYYY'))
    ) L1

So far so good. Now when I nest this in a subquery, and add a row_number() function, all my business_day values become 0. If I remove the row_number() function, business_day goes back to normal.

SELECT L2.DATES
, L2.business_day
, row_number() OVER (PARTITION BY L2.business_day ORDER BY L2.DATES ASC) as dateindex
FROM (
    SELECT DATES
        ,case when to_char(DATES, 'd') in (1,7)
         then 0
        else 1 end as business_day
    FROM (
        SELECT to_date('1/1/2020','MM/DD/YYYY') + (LEVEL -1) AS DATES
        FROM DUAL connect by level <=(to_date('1/1/2021','MM/DD/YYYY') - to_date('1/1/2020','MM/DD/YYYY'))
    ) L1
) L2

Any idea how adding a new column causes another's values to change?

Answer 1

I suspect you aren't paying attention to the actual dates; running your code in this db<>fiddle, the first query returns:

DATES     | BUSINESS_DAY
:-------- | -----------:
01-JAN-20 |            1
02-JAN-20 |            1
03-JAN-20 |            1
04-JAN-20 |            1
05-JAN-20 |            0
06-JAN-20 |            0
07-JAN-20 |            1
08-JAN-20 |            1
09-JAN-20 |            1
10-JAN-20 |            1
11-JAN-20 |            1
...

while the second returns:

DATES     | BUSINESS_DAY | DATEINDEX
:-------- | -----------: | --------:
05-JAN-20 |            0 |         1
06-JAN-20 |            0 |         2
12-JAN-20 |            0 |         3
13-JAN-20 |            0 |         4
19-JAN-20 |            0 |         5
20-JAN-20 |            0 |         6
26-JAN-20 |            0 |         7
27-JAN-20 |            0 |         8
02-FEB-20 |            0 |         9
03-FEB-20 |            0 |        10
...

All the business_day values are indeed zero... or at least, if you only look at the start of the result set. If you look further down:

...
27-DEC-20 |            0 |       103
28-DEC-20 |            0 |       104
01-JAN-20 |            1 |         1
02-JAN-20 |            1 |         2
03-JAN-20 |            1 |         3
04-JAN-20 |            1 |         4
...

You don't have an order-by clause, and the analytic processing internally happens to return in an order you aren't expecting. If you add an order-by then it looks more sensible, as in this db<>fiddle:

DATES     | BUSINESS_DAY | DATEINDEX
:-------- | -----------: | --------:
01-JAN-20 |            1 |         1
02-JAN-20 |            1 |         2
03-JAN-20 |            1 |         3
04-JAN-20 |            1 |         4
05-JAN-20 |            0 |         1
06-JAN-20 |            0 |         2
07-JAN-20 |            1 |         5
08-JAN-20 |            1 |         6
09-JAN-20 |            1 |         7
10-JAN-20 |            1 |         8
11-JAN-20 |            1 |         9
12-JAN-20 |            0 |         3
...

---

Incidentally, the 'd' format element is NLS-sensitive, so someone else running this code in a session with different settings could see different results. It would safer to do:

when to_char(DATES, 'Dy', 'NLS_DATE_LANGUAGE=ENGLISH') in ('Sat', 'Sun')