Special SQL windows function that works like a loop - sql

Question

I’m looking for some kind of SQL window function that calculate values based on a calculated value from a previous iteration when looping over the window. I’m not looking for ‘lag’ which will just take the original value of the previous row.
Here is the case: We have web analytics sessions. We would like to attribute to each session to the last relevant channel. There are 3 channels: direct, organic and paid. However, they have different priorities: paid will always be relevant. Organic will only be relevant if there was no paid channel in the last 30 days and direct would only be relevant if there was not paid or organic channel in the last 30 days
So in the example table we would like to calculate the values in column ‘attributed’ based on channel and the date columns. Note, the data is there for several users so this should be calculated per user.
+─────────────+───────+──────────+─────────────+
| date | user | channel | attributed |
+─────────────+───────+──────────+─────────────+
| 2022-01-01 | 123 | direct | direct |
| 2022-01-14 | 123 | paid | paid |
| 2022-02-01 | 123 | direct | paid |
| 2022-02-12 | 123 | direct | paid |
| 2022-02-13 | 123 | organic | paid |
| 2022-03-08 | 123 | direct | direct |
| 2022-03-10 | 123 | paid | paid |
+─────────────+───────+──────────+─────────────+
So in the table above row 1 is attributed direct because it’s the first line. The second then is paid as this has priority to direct. It stays paid for the next 2 sessions as direct has lower priority, then it switches to organic as the paid attribution is older than 30 days. The last one is then paid again as it has a higher priority than organic.
I would know how to solve it if you could decide whether a new channel needs to be attributed only based on the current row and the previous. I added here the SQL to do it:
with source as ( -- example data
select cast("2022-01-01" as date) as date, 123 as user, "direct" as channel
union all
select "2022-01-14", 123, "paid"
union all
select "2022-02-01", 123, "direct"
union all
select "2022-02-12", 123, "direct"
union all
select "2022-02-13", 123, "organic"
union all
select "2022-03-08", 123, "direct"
union all
select "2022-03-10", 123, "paid"
),
flag_new_channel as( -- flag sessions that would override channel informaton ; this only works statically here
select *,
case
when lag(channel) over (partition by user order by date) is null then 1
when date_diff(date,lag(date) over (partition by user order by date),day)>30 then 1
when channel = "paid" then 1
when channel = "organic" and lag(channel) over (partition by user order by date)!='paid' then 1
else 0
end flag
from source
qualify flag=1
)
select s.*,
f.channel attributed_channel,
row_number() over (partition by s.user, s.date order by f.date desc) rn -- number of flagged previous sessions
from source s
left join flag_new_channel f
on s.date>=f.date
qualify rn=1 --only keep the last flagged session at or before the current session
However, this would for example set "organic" in row 5 because it doesn't know "paid" is still relevant.
+─────────────+───────+──────────+─────────────────────+
| date | user | channel | attributed_channel |
+─────────────+───────+──────────+─────────────────────+
| 2022-01-01 | 123 | direct | direct |
| 2022-01-14 | 123 | paid | paid |
| 2022-02-01 | 123 | direct | paid |
| 2022-02-12 | 123 | direct | paid |
| 2022-02-13 | 123 | organic | organic |
| 2022-03-08 | 123 | direct | organic |
| 2022-03-10 | 123 | paid | paid |
+─────────────+───────+──────────+─────────────────────+
Any ideas? Not sure recursive queries can help or udfs. I’m using BigQuery usually but if you know solutions in other dialects it would still be interesting to know.

Answer 1

Here's one approach:
Updated: Corrected. I lost track of your correct / expected result, due to the confusing story.
For PostgreSQL, we can do something like this (with CTE and window functions):
The fiddle for PG 14
pri - provides a table of (channel, priority) pairs
cte0 - provides the test data
cte1 - determines the minimum priority over the last 30 days per user
final - the final query expression obtains the attributed channel name
WITH pri (channel, pri) AS (
VALUES ('paid' , 1)
, ('organic' , 2)
, ('direct' , 3)
)
, cte0 (date, xuser, channel) AS (
VALUES
('2022-01-01'::date, 123, 'direct')
, ('2022-01-14' , 123, 'paid')
, ('2022-02-01' , 123, 'direct')
, ('2022-02-12' , 123, 'direct')
, ('2022-02-13' , 123, 'organic')
, ('2022-03-08' , 123, 'direct')
, ('2022-03-10' , 123, 'paid')
)
, cte1 AS (
SELECT cte0.*
, pri.pri
, MIN(pri) OVER (PARTITION BY xuser ORDER BY date
RANGE BETWEEN INTERVAL '30' DAY PRECEDING AND CURRENT ROW
) AS mpri
FROM cte0
JOIN pri
ON pri.channel = cte0.channel
)
SELECT cte1.*
, pri.channel AS attributed
FROM cte1
JOIN pri
ON pri.pri = cte1.mpri
;
The result:
date
xuser
channel
pri
mpri
attributed
2022-01-01
123
direct
3
3
direct
2022-01-14
123
paid
1
1
paid
2022-02-01
123
direct
3
1
paid
2022-02-12
123
direct
3
1
paid
2022-02-13
123
organic
2
1
paid
2022-03-08
123
direct
3
2
organic
2022-03-10
123
paid
1
1
paid