Counting unique combinations up until a date - per month - sql

Question

I am looking into a table with transaction data of a two-sided platform, where you have buyers and sellers. I want to know the total amount of unique combinations of buyers and sellers. Let's say, Abe buys from Brandon in January, that's 1 combination. If Abe buys with Cece in February, that makes 2, but if Abe then buys from Brandon again, it's still 2.
My solution was to use the DENSE_RANK() function:
WITH
combos AS (
SELECT
t.buyerid, t.sellerid,
DENSE_RANK() OVER (ORDER BY t.buyerid, t.sellerid) AS combinations
FROM transactions t
WHERE t.transaction_date < '2018-05-01'
)
SELECT
MAX(combinations) AS total_combinations
FROM combos
This works fine. Each new combo gets a higher rank, and if you select the MAX of that result, you know the amount of unique combos.
However, I want to know this total amount of unique combos on a per month basis. The problem here is that if I group per transaction month, it only counts the unique combos in that month. In the example of Abe, it would be a unique combo in January, and then another combo in the next month, because that's how grouping works in SQL.
Example:
transaction_date buyerid sellerid
2018-01-03 3828 219
2018-01-08 2831 123
2018-02-10 3828 219
The output of DENSE_RANK() named combinations over all these rows is:
transaction_date buyerid sellerid combinations
2018-01-03 3828 219 1
2018-01-08 2831 123 2
2018-02-10 3828 219 2
And therefore, when selecting the MAX combinations you know the amount of unique buyer/seller combos, which is here.
However, I would like to see a running total of unique combos up until each start of the month, for all months until now. But, when we group on month, it would go like this:
transaction_date buyerid sellerid month combinations
2018-01-03 3828 219 jan 1
2018-01-08 2831 123 jan 2
2018-02-10 3828 219 feb 1
While I actually would want an output like:
month total_combinations_at_month_start
jan 0
feb 2
mar 2
How should I solve this? I've tried to find help on all kinds of window functions, but no luck until now. Thanks!

Answer 1

Here is one method:
WITH combos AS (
SELECT t.*,
ROW_NUMBER() OVER (PARTITION BY sellerid, buyerid ORDER BY t.transaction_date) as combo_seqnum,
ROW_NUMBER() OVER (PARTITION BY sellerid, buyerid, date_trunc('month', t.transaction_date) ORDER BY t.transaction_date) as combo_month_seqnum
FROM transactions t
WHERE t.transaction_date < '2018-05-01'
)
SELECT 'Overall' as which, COUNT(*)
FROM combos
WHERE combo_seqnum = 1
UNION ALL
SELECT to_char(transaction_date, 'YYYY-MM'), COUNT(*)
FROM combos
WHERE combo_month_seqnum = 1
GROUP BY to_char(transaction_date, 'YYYY-MM');
This puts the results in separate rows. If you want a cumulative number and number per month:
SELECT to_char(transaction_date, 'YYYY-MM'),
SUM( (combo_month_seqnum = 1)::int ) as uniques_in_month,
SUM(SUM( (combo_seqnum = 1)::int )) OVER (ORDER BY to_char(transaction_date, 'YYYY-MM')) as uniques_through_month
FROM combos
GROUP BY to_char(transaction_date, 'YYYY-MM')
Here is a rextester illustrating the solution.