Postgresql running totals with groups missing data and outer joins - postgresql

Question

I've written a sql query that pulls data from a user table and produces a running total and cumulative total of when users were created. The data is grouped by week (using the windowing feature of postgres). I'm using a left outer join to include the weeks when no users where created. Here is the query...
<!-- language: lang-sql -->
WITH reporting_period AS (
SELECT generate_series(date_trunc('week', date '2015-04-02'), date_trunc('week', date '2015-10-02'), interval '1 week') AS interval
)
SELECT
date(interval) AS interval
, count(users.created_at) as interval_count
, sum(count( users.created_at) ) OVER (order by date_trunc('week', users.created_at)) AS cumulative_count
FROM reporting_period
LEFT JOIN users
ON interval=date(date_trunc('week', users.created_at) )
GROUP BY interval, date_trunc('week', users.created_at) ORDER BY interval
It works almost perfectly. The cumulative value is calculated properly for weeks week a user was created. For weeks when no user was create it is set to grand total and not the cumulative total up to that point.
Notice the rows with ** the Week Tot column (interval_count) is 0 as expected but the Run Tot (cumulative_total) is 1053 which equals the grand total.
Week Week Tot Run Tot
-----------------------------------
2015-03-30 | 4 | 4
2015-04-06 | 13 | 17
2015-04-13 | 0 | 1053 **
2015-04-20 | 9 | 26
2015-04-27 | 3 | 29
2015-05-04 | 0 | 1053 **
2015-05-11 | 0 | 1053 **
2015-05-18 | 1 | 30
2015-05-25 | 0 | 1053 **
...
2015-06-08 | 996 | 1031
...
2015-09-07 | 2 | 1052
2015-09-14 | 0 | 1053 **
2015-09-21 | 1 | 1053 **
2015-09-28 | 0 | 1053 **
This is what I would like
Week Week Tot Run Tot
-----------------------------------
2015-03-30 | 4 | 4
2015-04-06 | 13 | 17
2015-04-13 | 0 | 17 **
2015-04-20 | 9 | 26
2015-04-27 | 3 | 29
2015-05-04 | 0 | 29 **
...
It seems to me that if the outer join can somehow apply the grand total to the last column it should be possible to apply the current running total but I'm at a loss on how to do it.
Is this possible?

Answer 1

This is not guaranteed to work out of the box as I havent tested on acutal tables, but the key here is to join users on created_at over a range of dates.
with reportingperiod as (
select intervaldate as interval_begin,
intervaldate + interval '1 month' as interval_end
from (
SELECT GENERATE_SERIES(DATE(DATE_TRUNC('day', DATE '2015-03-15')),
DATE(DATE_TRUNC('day', DATE '2015-10-15')), interval '1 month') AS intervaldate
) as rp
)
select interval_end,
interval_count,
sum(interval_count) over (order by interval_end) as running_sum
from (
select interval_end,
count(u.created_at) as interval_count
from reportingperiod rp
left join (
select created_at
from users
where created_at < '2015-10-02'
) u on u.created_at > rp.interval_begin
and u.created_at <= rp.interval_end
group by interval_end
) q

Answer 2

I figured it out. The trick was subqueries. Here's my approach
Add a count column to the generate_series call with default value of 0
Select interval and count(users.created_at) from the users data
Union the the generate_series and the result from the select in step #2
(At this point the result will have duplicates for each interval)
Use the results in a subquery to get interval and max(interval_count) which eliminates duplicates
Use the window aggregation as before to get the running total
SELECT
interval
, interval_count
, SUM(interval_count ) OVER (ORDER BY interval) AS cumulative_count
FROM
(
SELECT interval, MAX(interval_count) AS interval_count FROM
(
SELECT GENERATE_SERIES(DATE(DATE_TRUNC('week', DATE '2015-04-02')),
DATE(DATE_TRUNC('week', DATE '2015-10-02')), interval '1 week') AS interval,
0 AS interval_count
UNION
SELECT DATE_TRUNC('week', users.created_at) AS INTERVAL,
COUNT(users.created_at) AS interval_count FROM users
WHERE users.created_at < date '2015-10-02'
GROUP BY 1 ORDER BY 1
) sub1
GROUP BY interval
) grouped_data
I'm not sure if there are any serious performance issues with this approach but it seems to work. If anyone has a better, more elegant or performant approach I would love the feedback.
Edit: My solution doesn't work when trying to group by arbitrary time windows
Just tried this solution with the following changes
/* generate series using DATE_TRUNC('day'...)*/
SELECT GENERATE_SERIES(DATE(DATE_TRUNC('day', DATE '2015-04-02')),
DATE(DATE_TRUNC('day', DATE '2015-10-02')), interval '1 month') AS interval,
0 AS interval_count
/* And this part */
SELECT DATE_TRUNC('day', users.created_at) AS INTERVAL,
COUNT(users.created_at) AS interval_count FROM users
WHERE users.created_at < date '2015-10-02'
GROUP BY 1 ORDER BY 1
For example is is possible to produce these similar results but have the data grouped by intervals as so
3/15/15 - 4/14/15,
4/15/15 - 5/14/15,
5/15/15 - 6/14/15
etc.