Number of unique dates - sql

Question

There is table:
CREATE TABLE my_table
(gr_id NUMBER,
start_date DATE,
end_date DATE);
All dates always have zero time portion. I need to know a fastest way to compute number of unique dates inside gr_id.
For example, if there is rows (dd.mm.rrrr):
1 | 01.01.2000 | 07.01.2000
1 | 01.01.2000 | 07.01.2000
2 | 01.01.2000 | 03.01.2000
2 | 05.01.2000 | 07.01.2000
3 | 01.01.2000 | 04.01.2000
3 | 03.01.2000 | 05.01.2000
then right answer will be
1 | 7
2 | 6
3 | 5
At now I use additional table
CREATE TABLE mfr_date_list
(MFR_DATE DATE);
with every date between 01.01.2000 and 31.12.2020 and query like this:
SELECT COUNT(DISTINCT mfr_date_list.mfr_date) cnt,
dt.gr_id
FROM dwh_mfr.mfr_date_list,
(SELECT gr_id,
start_date AS sd,
end_date AS ed
FROM my_table
) dt
WHERE mfr_date_list.mfr_date BETWEEN dt.sd AND dt.ed
AND dt.ed IS NOT NULL
GROUP BY dt.gr_id
This query return correct resul data set, but I think it's not fastest way. I think there is some way to build query withot table mfr_date_list at all.
Oracle 11.2 64-bit.

Answer 1

I would expect what you're doing to be the fastest way (as always test). Your query can be simplified, though this only aids understanding and not necessarily speed:
select t.gr_id, count(distinct dl.mfr_date) as cnt
from my_table t
join mfr_date_list dl
on dl.mfr_date between t.date_start and t.date_end
where t.end_date is not null
group by t.gr_id
Whatever you do you have to generate the data between the two dates somehow as you need to remove the overlap. One way would be to use CAST(MULTISET()), as Lalit Kumar explains:
select gr_id, count(distinct end_date - column_value + 1)
from my_table m
cross join table(cast(multiset(select level
from dual
connect by level <= m.end_date - m.start_date + 1
) as sys.odcinumberlist))
group by gr_id;
GR_ID COUNT(DISTINCTEND_DATE-COLUMN_VALUE+1)
---------- --------------------------------------
1 7
2 6
3 5
This is very Oracle specific but should perform substantially better than most other row-generators as you're only accessing the table once and you're generating the minimal number of rows required due to the condition linking MY_TABLE and your generated rows.

Answer 2

What you really need to do is combine the ranges and then count the lengths. This can be quite challenging because of duplicate dates. The following is one way to approach this.
First, enumerate the dates and determine whether the date is "in" or "out". When the cumulative sum is 0 then it is "out":
select t.gr_id, dt,
sum(inc) over (partition by t.gr_id order by dt) as cume_inc
from (select t.gr_id, t.start_date as dt, 1 as inc
from my_table t
union all
select t.gr_id, t.end_date + 1, -1 as inc
from my_table t
) t
Then, use lead() to determine how long the period is:
with inc as (
select t.gr_id, dt,
sum(inc) over (partition by t.gr_id order by dt) as cume_inc
from (select t.gr_id, t.start_date as dt, 1 as inc
from my_table t
union all
select t.gr_id, t.end_date + 1, -1 as inc
from my_table t
) t
)
select t.gr_id,
sum(nextdt - dt) as daysInUse
from (select inc.*, lead(dt) over (partition by t.gr_id order by dt) as nextdt
from inc
) t
group by t.gr_id;
This is close to what you want. The following are two challenges: (1) putting in the limits and (2) handling ties. The following should work (although there might be off-by-one and boundary issues):
with inc as (
select t.gr_id, dt, priority,
sum(inc) over (partition by t.gr_id order by dt) as cume_inc
from ((select t.gr_id, t.start_date as dt, count(*) as inc, 1 as priority
from my_table t
group by t.gr_id, t.start_date
)
union all
(select t.gr_id, t.end_date + 1, - count(*) as inc, -1
from my_table t
group by t.gr_id, t.end_date
)
) t
)
select t.gr_id,
sum(least(nextdt, date '2020-12-31') - greatest(dt, date, '2010-01-01')) as daysInUse
from (select inc.*, lead(dt) over (partition by t.gr_id order by dt, priority) as nextdt
from inc
) t
group by t.gr_id;