Computing rolling sums efficiently in PostgreSQL - postgresql

Question

Supposing I have a set of transactions (purchases) with dates for a set of customers, I want to calculate a rolling x day sum of purchase amount and number of purchases by customer in that same window. I've gotten it to work using a window function, but I have to fill in for dates where the customer did not make any purchases. In so doing, I'm using a Cartesian product. Is there a more efficient approach so that it's more scalable as the number of customers – and time window – increases?
Edit: As noted in the comments, I'm on PostgreSQL v9.3.
Here's sample data (note that some customers may have 0, 1, or multiple purchases on a given date):
| id | cust_id | txn_date | amount |
|----|---------|------------|--------|
| 1 | 123 | 2017-08-17 | 10 |
| 2 | 123 | 2017-08-17 | 5 |
| 3 | 123 | 2017-08-18 | 5 |
| 4 | 123 | 2017-08-20 | 50 |
| 5 | 123 | 2017-08-21 | 100 |
| 6 | 456 | 2017-08-01 | 5 |
| 7 | 456 | 2017-08-01 | 5 |
| 8 | 456 | 2017-08-01 | 5 |
| 9 | 456 | 2017-08-30 | 5 |
| 10 | 456 | 2017-08-01 | 1000 |
| 11 | 789 | 2017-08-15 | 1000 |
| 12 | 789 | 2017-08-30 | 1000 |
And here's the desired output:
| cust_id | txn_date | sum_dly_txns | tot_txns_7d | cnt_txns_7d |
|---------|------------|--------------|-------------|-------------|
| 123 | 2017-08-17 | 15 | 15 | 2 |
| 123 | 2017-08-18 | 5 | 20 | 3 |
| 123 | 2017-08-20 | 50 | 70 | 4 |
| 123 | 2017-08-21 | 100 | 170 | 5 |
| 456 | 2017-08-01 | 1015 | 1015 | 4 |
| 456 | 2017-08-30 | 5 | 5 | 1 |
| 789 | 2017-08-15 | 1000 | 1000 | 1 |
| 789 | 2017-08-30 | 1000 | 1000 | 1 |
Here's SQL that produces the totals as desired:
SELECT *
FROM (
-- One row per day per user
WITH daily_txns AS (
SELECT
t.cust_id
,t.txn_date AS txn_date
,SUM(t.amount) AS sum_dly_txns
,COUNT(t.id) AS cnt_dly_txns
FROM transactions t
GROUP BY t.cust_id, txn_date
),
-- Every possible transaction date for every user
dummydates AS (
SELECT txn_date, uids.cust_id
FROM (
SELECT generate_series(
timestamp '2017-08-01'
,timestamp '2017-08-30'
,interval '1 day')::date
) d(txn_date)
CROSS JOIN (SELECT DISTINCT cust_id FROM daily_txns) uids
),
txns_dummied AS (
SELECT
d.cust_id
,d.txn_date
,COALESCE(sum_dly_txns,0) AS sum_dly_txns
,COALESCE(cnt_dly_txns,0) AS cnt_dly_txns
FROM dummydates d
LEFT JOIN daily_txns dx
ON d.txn_date = dx.txn_date
AND d.cust_id = dx.cust_id
ORDER BY d.txn_date, d.cust_id
)
SELECT
cust_id
,txn_date
,sum_dly_txns
,SUM(COALESCE(sum_dly_txns,0)) OVER w AS tot_txns_7d
,SUM(cnt_dly_txns) OVER w AS cnt_txns_7d
FROM txns_dummied
WINDOW w AS (
PARTITION BY cust_id
ORDER BY txn_date
ROWS BETWEEN 6 PRECEDING AND CURRENT ROW -- 7d moving window
)
ORDER BY cust_id, txn_date
) xfers
WHERE sum_dly_txns > 0 -- Omit dates with no transactions
;
SQL Fiddle

Answer 1

Instead of ROWS BETWEEN 6 PRECEDING AND CURRENT ROW did you want to write RANGE '6 days' PRECEEDING ?
This must be what you are looking for:
SELECT DISTINCT
cust_id
,txn_date
,SUM(amount) OVER (PARTITION BY cust_id, txn_date) sum_dly_txns
,SUM(amount) OVER (PARTITION BY cust_id ORDER BY txn_date RANGE '6 days' PRECEDING)
,COUNT(*) OVER (PARTITION BY cust_id ORDER BY txn_date RANGE '6 days' PRECEDING)
from transactions
ORDER BY cust_id, txn_date
Edit: Since you are using an old version (I tested the one above on my postgresql 11), the point above will not make much sense so you will need to old-fashioned SQL (that is, witout window functions).
It is a bit less efficient but does a fair job.
WITH daily_txns AS (
SELECT
t.cust_id
,t.txn_date AS txn_date
,SUM(t.amount) AS sum_dly_txns
,COUNT(t.id) AS cnt_dly_txns
FROM transactions t
GROUP BY t.cust_id, txn_date
)
SELECT t1.cust_id, t1.txn_date, t1.sum_dly_txns, SUM(t2.sum_dly_txns), SUM(t2.cnt_dly_txns)
from daily_txns t1
join daily_txns t2 ON t1.cust_id = t2.cust_id and t2.txn_date BETWEEN t1.txn_date - 7 and t1.txn_date
group by t1.cust_id, t1.txn_date, t1.sum_dly_txns
order by t1.cust_id, t1.txn_date