Get daily MAU rolling sum using PostgreSQL - postgresql

Question

I have daily log data stored in a Postgres database structured with an id and date. Users can, obviously, have multiple rows in the database if they log in multiple times.
To visualize:
| id | timestamp |
|------|---------------------|
| 0099 | 2004-10-19 10:23:54 |
| 1029 | 2004-10-01 10:23:54 |
| 2353 | 2004-10-20 8:23:54 |
Let's say MAU ("monthly active users") is defined as the number of unique ids that log in for a given calendar month. I would like to get the rolling sum of MAU for each day in a month, i.e. MAU at different points in time as it grows. For example, if we were looking at October 2014:
| date | MAU |
|------------|-------|
| 2014-10-01 | 10000 |
| 2014-10-02 | 12948 |
| 2014-10-03 | 13465 |
And so forth until the end of the month. I've heard that window functions might be one way to solve this. Any ideas how to utilize that to get the rolling MAU sum?

Answer 1

After reading the documentation for Postgres window functions, here's one solution that gets the rolling MAU sum for the current month:
-- First, get id and date of each timestamp within the current month
WITH raw_data as (SELECT id, date_trunc('day', timestamp) as timestamp
FROM user_logs
WHERE date_trunc('month', timestamp) = date_trunc('month', current_timestamp)),
-- Since we only want to count the earliest login for a month
-- for a given login, use MIN() to aggregate
month_data as (SELECT id, MIN(timestamp) as timestamp_day FROM raw_data GROUP BY id)
-- Postgres doesn't support DISTINCT for window functions, so query
-- from the rolling sum to have each row as a day
SELECT timestamp_day as date, MAX(count) as MAU
FROM (SELECT timestamp_day, COUNT(id) OVER(ORDER BY timestamp_day) FROM month_data) foo
GROUP By timestamp_day

Answer 2

For a given month, you can calculate this by adding in a user on the first day during the month when they are seen:
select date_trunc('day', mints), count(*) as usersOnDay,
sum(count(*)) over (order by date_trunc('day', mints)) as cume_users
from (select id, min(timestamp) as mints
from log
where timestamp >= '2004-10-01'::date and timestamp < '2004-11-01'::date
group by id
) l
group by date_trunc('day', mints);
Note: This answers your question about one month. This can be extended to more calendar months, where you are counting the unique users on the first day and then adding increments.
If you have a question where the cumulative period passes month boundaries, then ask another question and explain what a month means under those circumstances.