SQL Window Function over sliding time window - sql

Question

I have the following data:
country objectid objectuse
record_date
2022-07-20 chile 0 4
2022-07-01 chile 1 4
2022-07-02 chile 1 4
2022-07-03 chile 1 4
2022-07-04 chile 1 4
... ... ... ...
2022-07-26 peru 3088 4
2022-07-27 peru 3088 4
2022-07-28 peru 3088 4
2022-07-30 peru 3088 4
2022-07-31 peru 3088 4
The data describes the daily usage of an object within a country for a single month (July 2022), and not all object are used every day. One of the things I am interested in finding is the sum of the monthly maximums for the month:
WITH month_max AS (
SELECT
country,
objectid,
MAX(objectuse) AS maxuse
FROM mytable
GROUP BY
country,
objectid
)
SELECT
country,
SUM(maxuse)
FROM month_max
GROUP BY country;
Which results in this:
country sum
-------------
chile 1224
peru 17008
But what I actually want is to get the rolling sum of the maxima from the beginning of the month up to each date. So that I get something that looks like:
country sum
record_date
2022-07-01 chile 1
2022-07-01 peru 1
2022-07-02 chile 2
2022-07-02 peru 3
... ... ...
2022-07-31 chile 1224
2022-07-31 peru 17008
I tried using a window function like this to no avail:
SELECT
*,
SUM(objectuse) OVER (
PARTITION BY country
ORDER BY record_date ROWS 30 PRECEDING
) as cumesum
FROM mytable
order BY cumesum DESC;
Is there a way I can achieve the desired result in SQL?
Thanks in advance.
EDIT: For what it's worth, I asked the same question but on Pandas and I received an answer; perhaps it helps to figure out how to do it in SQL.

Answer 1

What ended up working is probably not the most efficient approach to this problem. I essentially created backwards looking blocks from each day in the month back towards the beginning of the month. Within each of these buckets I get the maximum of objectuse for each objectid within that bucket. After taking the max, I sum across all the maxima for that backward looking period. I do this for every day in the data.
Here is the query that does it:
WITH daily_lookback AS (
SELECT
A.record_date,
A.country,
B.objectid,
MAX(B.objectuse) AS maxuse
FROM mytable AS A
LEFT JOIN mytable AS B
ON A.record_date >= B.record_date
AND A.country = B.country
AND DATE_PART('month', A.record_date) = DATE_PART('month', B.record_date)
AND DATE_PART('year', A.record_date) = DATE_PART('year', B.record_date)
GROUP BY
A.record_date,
A.country,
B.objectid
)
SELECT
record_date,
country,
SUM(maxuse) AS usetotal
FROM daily_lookback
GROUP BY
record_date,
country
ORDER BY
record_date;
Which gives me exactly what I was looking for: the cumulative sum of the objectid maximums for the backward looking period, like this:
country sum
record_date
2022-07-01 chile 1
2022-07-01 peru 1
2022-07-02 chile 2
2022-07-02 peru 3
... ... ...
2022-07-31 chile 1224
2022-07-31 peru 17008

Answer 2

You need to change your inner query to use the windowed maximum:
WITH month_max AS (
SELECT record_date, country, objectid,
MAX(objectuse) over (PARTITION BY country, objectid ORDER BY record_date) AS mx
FROM mytable
)
SELECT record_date, country, SUM(mx) as "sum"
FROM month_max
GROUP BY record_date, country;
This does assume one row per object per date.
Here's a re-written version of your query. With indexing it seems possible that it might run faster:
select record_date, country, min(usetotal) as usetotal
from mytable d inner join lateral (
select distinct sum(max(objectuse)) over () as usetotal from mytable a
where a.record_date between date_trunc('month', d.record_date) and d.record_date
and a.country = d.country
group by objectid
) T on 1 = 1
group by record_date, country
order by record_date, country;
https://dbfiddle.uk/?rdbms=postgres_14&fiddle=63760e30aecf4c885ec4967045b6cd03