if I analyze week 22 in 2020, an active customer will be the one that has bought at least once since week 22, 2019.
Problems on your side
This method has some corner case ambiguities / issues:
Do you include or exclude "week 22 in 2020"? (I exclude it below to stay closer to "a year".)
A year can have 52 or 53 full weeks. Depending on the current date, the calculation is based on 52 or 53 weeks, causing a possible bias of almost 2 %!
If you start the time range on "the same date last year", then the margin of error is only 1 / 365 or ~ 0.3 %, due to leap years.
A fixed "period of 365 days" (or 366) would eliminate the bias altogether.
Problems on the SQL side
Unfortunately, window functions do not currently allow the DISTINCT
key word (for good reasons). So something of the form:
SELECT count(DISTINCT email) OVER (ORDER BY year, week
GROUPS BETWEEN 52 PRECEDING AND 1 PRECEDING)
FROM ...
.. triggers:
ERROR: DISTINCT is not implemented for window functions
The GROUPS
keyword has only been added in Postgres 10 and would otherwise be just what we need.
What's more, your odd frame definition wouldn't even work exactly, since the number of weeks to consider is not always 52, as discussed above.
So we have to roll our own.
Solution
The following simply generates all weeks of interest, and computes the distinct count of customers for each. Simple, except that date math is never entirely simple. But, depending on details of your setup, there may be faster solutions. (I had several other ideas.)
The time range for which to report may change. Here is an auxiliary function to generate weeks of a given year:
CREATE OR REPLACE FUNCTION f_weeks_of_year(_year int)
RETURNS TABLE(year int, week int, week_start timestamp)
LANGUAGE sql IMMUTABLE STRICT PARALLEL SAFE
ROWS 52 COST 10 AS
$func$
SELECT _year, d.week::int, d.week_start
FROM generate_series(date_trunc('week', make_date(_year, 01, 04)::timestamp) -- first day of first week
, LEAST(date_trunc('week', localtimestamp), make_date(_year, 12, 28)::timestamp) -- latest possible start of week
, interval '1 week') WITH ORDINALITY d(week_start, week)
$func$;
Call:
SELECT * FROM f_weeks_of_year(2020);
It returns 1 row per week, but stops at the current week for the current year. (Empty set for future years.)
The calculation is based on these facts:
- The first ISO week of the year always contains January 04.
- The last ISO week cannot start after December 28.
Actual week numbers are computed on the fly using WITH ORDINALITY
. See:
Aside, I stick to timestamp
and avoid timestamptz
for this purpose. See:
The function also returns the timestamp of the start of the week (week_start
), which we don't need for the problem at hand. But I left it in to make the function more useful in general.
Makes the main query simpler:
WITH weekly_customer AS (
SELECT DISTINCT
EXTRACT(YEAR FROM orderdate)::int AS year
, EXTRACT(WEEK FROM orderdate)::int AS week
, email
FROM orders_table
WHERE paid
AND orderdate >= date_trunc('week', timestamp '2019-01-04') -- max range for 2020!
ORDER BY 1, 2, 3 -- optional, might improve performance
)
SELECT d.year, d.week
, (SELECT count(DISTINCT email)
FROM weekly_customer w
WHERE (w.year, w.week) >= (d.year - 1, d.week) -- row values, see below
AND (w.year, w.week) < (d.year , d.week) -- exclude current week
) AS active_customers
FROM f_weeks_of_year(2020) d; -- (year int, week int, week_start timestamp)
db<>fiddle here
The CTE weekly_customer
folds to unique customers per calendar week once, as duplicate entries are just noise for our calculation. It's used many times in the main query. The cut-off condition is based on Jan 04 once more. Adjust to your actual reporting period.
The actual count is done with a lowly correlated subquery. Could be a LEFT JOIN LATERAL ... ON true
instead. See:
Using row value comparison to make the range definition simple. See: