How to select and group fortnightly in postgreql - sql

Question

I am trying to group the rows in a table fortnightly, but can't seem to work out how to do it - especially, as the date_part function does not have a 'fortnight' keyword argument.
This is what I have so far:
CREATE TABLE foo(
dt DATE NOT NULL,
f1 REAL NOT NULL,
f2 REAL NOT NULL,
f3 REAL NOT NULL,
f4 REAL NOT NULL
);
SELECT AVG((f1+f2+f3+f4)/4) as fld_avg FROM
(
SELECT date_part('year', dt) AS year_part,
date_part('fortnight', dt) AS fortnight_part,
f1, f2, f3, f4
FROM foo
WHERE dt >= date_trunc('day', NOW() - '3 month')
) foo
GROUP BY year_part, fortnight_part
How may I rewrite (or modify) the query above so as to group data fortnightly?

Answer 1

Basic idea
What we need to do, is take intervals of 14 consecutive days and map them to unique buckets and then group by those buckets. These buckets can of any type, int, char, timstamp, as long as we have unique value.
Division
A simple way to accomplish this is division. Divide by 14 days and truncate the result to date precision.
For example, we can extract the number of seconds since 1970-01-01, the UNIX epoch, and divide by the number of seconds in a fortnight: 14 * 24 * 60 * 60 = 14 * 86400 = 1209600. (I'll use Vao Tsun's example data)
WITH c(d) AS (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
SELECT (EXTRACT(EPOCH FROM d)::int/86400)/14 fortnight FROM c
which yields fortnights since 1970-01-01 (a Thursday):
fortnight
-----------
1251
1252
1254
1254
(4 rows)
The integer values we get, represent the number of fortnights since 1970-01-01, but we don't have to care about this. The important thing is, that it uniquely identifies a fortnight.
Due to 1970-01-01 being a Thursday, all fortnights will start at a Thursday. We might want to vary the starting point of our fortnight to a different day of the week (e.g. Monday) by adding:
WITH c(d) AS (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
SELECT (EXTRACT(EPOCH FROM d)::int/86400 + 4)/14 fortnight FROM c
By adding four days to Thursday we end up at Monday.
If you rather want fortnights with respect to the beginning of the year, instead of some arbitrary absolute date, such as 1970-01-01, we can use the day of the year instead:
WITH c(d) AS (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
SELECT EXTRACT(year FROM d) * 26 + EXTRACT(doy FROM d)::int/14 AS fortnight FROM c;
which yields
fortnight
-----------
52467
52468
52469
52470
(4 rows)
We need to multiply the extracted year by 26, because there are 26.1… fortnights in a year.
Truncation
Instead of division another approach is truncation. We map each day of a specific fortnight to the first timestamp of that fortnight.
WITH c(d) AS (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
SELECT d - make_interval(secs => EXTRACT(EPOCH FROM d)::int % (86400 * 14)) AS fortnight FROM c;
which yields
fortnight
---------------------
2017-12-14 00:00:00
2017-12-28 00:00:00
2018-01-25 00:00:00
2018-01-25 00:00:00
(4 rows)
This might seems a bit more complicated, but has some benefits. The result is still a date/time type and other code does not need to worry about the fact, that we used fortnights.
Again, instead of absolute fortnights, we can calculate this with respect to the beginning of the year:
WITH c(d) AS (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
SELECT d - make_interval(days => EXTRACT(dow FROM d)::int % 14) AS fortnight FROM c;
which yields
fortnight
---------------------
2017-12-17 00:00:00
2017-12-31 00:00:00
2018-01-21 00:00:00
2018-01-28 00:00:00
(4 rows)
The result is of type timestamp, you might want to have date instead. This can be addressed by casting:
(d - make_interval(days => EXTRACT(dow FROM d)::int % 14))::date
or subtracting int instead of interval from date:
d - (EXTRACT(dow FROM d)::int % 14)
There are much more possibilities. With this scheme, we can calculate the fortnight or any other interval with respect to the beginning of the month, some arbitrary date, etc.

Answer 2

update
fortnight is a two week period - one even the other odd. eg week 1 and 2, 3 and 4, 5 and 6.
closer: 2 is even, mod(2,2)=0 and 1 is odd, mod(1,2)=1
4 is even, mod(4,2)=0 and 3 is odd, mod(3,2)=1
6 is even, mod(6,2)=0 and 5 is odd, mod(5,2)=1
thus you can make an assumption that each one week's in year consecutive number divided by two reminder is 1, and each next one weeks number/2 reminders 0
The general idea is - using the sequential number of week in a year. To avoid Jan 1st to be first and Dec31 (possible be the 53rd - and thus two odds in a row), I use IW
week number of ISO 8601 week-numbering year (01-53; the first Thursday
of the year is in week 1)
then I assume that if one week number will be odd, next will be even, so we divide all the time in parts of two weeks - even+odd.
SQL Example:
o=# with c(d) as (values('2017.12.21'::date),('2017.12.31'),('2018.01.26'),('2018.02.01'))
select d,to_char(d,'IW'),right(to_char(d,'IW'),1)::int,mod(right(to_char(d,'IW'),1)::int, 2) from c;
d | to_char | right | mod
------------+---------+-------+-----
2017-12-21 | 51 | 1 | 1
2017-12-31 | 52 | 2 | 0
2018-01-26 | 04 | 4 | 0
2018-02-01 | 05 | 5 | 1
(4 rows)
mod is either 0 or 1 - group by this column
https://www.postgresql.org/docs/current/static/functions-math.html
https://www.postgresql.org/docs/current/static/functions-formatting.html
Of course you would need to add outer join on generate_series if you want data without gaps...

Answer 3

I post another answer to explain how I was wrong and why my "smart-n-neat"
way failed...
the schema build and queries are at:
https://www.db-fiddle.com/f/j5i2Td8CvxCVXQQYePKzCe/0
the first (and correct) query:
select distinct w2, avg(c) over (partition by w2)
from d
join generate_series('2016.11.28'::date,'2017.02.23'::date,'2 weeks'::interval) w2
on gs >= w2 and gs < w2 + '2 weeks'::interval
order by w2;
Is a long, simple and correct approach. with idea is to join on two weeks interval. It's working, reliable and all good.
Now the second query:
select distinct div(to_char(gs,'IW')::int,2), min(gs) over w, avg(c) over w
from d
window w as (partition by div(to_char(gs,'IW')::int,2))
order by min;
Is much shorter, neater and smarter, yet has a huge limitation and is unusable. Here's why:
My approach splits next to last two-weeks-interval to two parts: last week of 2016 and first week of 2017, thus dividing the result by half. If you multiply a sum of averages for those two weeks by a half, the result for both queries will match. Alas introducing CASE WHEN logic for the edge year weeks makes neat solution a heavy and overhead. And thus the very point is lost.
TL;DR the neat and lightweight solution works only on interval of one year, farther then two weeks from end or start of the year and lastly if our fortnightly interval starts from Monday.
Now the idea behind lightweight solution: round(2/2, 0)=1 and round(3/2, 0)=1, so you can divide year in intervals of two weeks and use it for grouping by.
Also I deliberately took not this New Year switch, because this 2018 Jan 1 is Monday, so IW is same as WW - which usually is not the case.
Lastly my first answer with odd and even weeks is not viable at all. It divides year not in two-weeks interval, but rather in two parts - for even and odd weeks... I deceived myself with "something close" idea and worked on the reminder, while I should do the opposite the whole value of division...