Using windows functions to count by groups of dates PostgreSQL - sql

Question

I have a table amongst whose columns are id and created_at and I want to use window functions around the created_at of each entry to count how many entries there are within 48 hours of them. As an example, for the original table:
id | created_at
----|------------
01 | 2016/01/04
02 | 2016/01/05
03 | 2016/01/05
04 | 2016/01/06
05 | 2016/01/07
06 | 2016/01/08
07 | 2016/01/08
08 | 2016/01/09
and the result should be
id | created_at | count
----|------------|-------
01 | 2016/01/04 | 4
02 | 2016/01/05 | 5
03 | 2016/01/05 | 5
04 | 2016/01/06 | 7
05 | 2016/01/07 | 7
06 | 2016/01/08 | 5
07 | 2016/01/08 | 5
08 | 2016/01/09 | 4
The explanation is that since there are 2 transactions on 2016/01/05, 1 on 2016/01/06, 1 on 2016/01/07, 2 on 2016/01/08, and 1 on 2016/01/09, there are a total of 7 transactions within 2 days of transaction 05.

Answer 1

It is better to use a date table that have consecutive dates in case dates in your table have gaps.
I am wondering what's the role of the id column? Here is how I would do it without considering the id column.
select row_number()over(order by dt) as id
,dt as created_at
,cnt1+cnt2+cnt3+cnt4+cnt5 as cnt
from
(
select
date_table.dt
,lag(cnt,2,0)over(order by created_at asc) as cnt1
,lag(cnt,1,0)over(order by created_at asc) as cnt2
,isnull(cnt,0) cnt3
,lead(cnt,1,0)over(order by created_at asc) as cnt4
,lead(cnt,2,0)over(order by created_at asc) as cnt5
from
date_table left join
(select created_at,count(*) as cnt from your_table group by created_at) c
on date_table.day = c.created_at
) T

Answer 2

Using window functions for this purpose is challenging because of the duplicate days. You can get the results using a join or correlated subquery:
select t.*,
(select count(*)
from t t2
where t2 between t.created_at - interval 2 * '1 day' and
t.created_at + interval 2 * '1 day'
) as cnt
from t;
EDIT:
You could use window functions by doing a cumulative sum by date and then joining back. This is, of course, a bit challenging because of holes in the dates. But, something like this:
with c as (
select d.dte, count(t.created_at) as cnt,
sum(count(t.created_at))) over (order by d.dte) as cumecnt
from (select generate_series(min(created_at) - interval '2 day',
max(created_at) + interval '2 day',
'1 day')
from t
) d(dte) left join
on d.dte = t.created_at
)
select t.*, cmax.cumecnt - cmin.cumecnt
from t join
c cmin
on t.created_at = cmin.dte + interval '2 day' join
c cmax
on t.created_at = cmax.dte - interval '2 day';