Forward fill since (possibly non existent) date in BigQuery - google-bigquery

Question

I have data from two different sources. On one hand I have user data from our app. This has a primary key of ID and UTC date. There are only rows for UTC dates when are users uses the app. On the other hand I have advertisement campaign attribition data for the users (which can be multiple advertisment campaigns per user). This table has a primary key of ID and campaign and a metric containing a advertisment attribution timestamp. I want to combine the two data sources such that I can compute if a campaign is generating more revenue than it costs among other campaign statistics.
App data example:
SELECT
*
FROM UNNEST(ARRAY<STRUCT<ID INT64, UTC_Date DATE, Revenue FLOAT64>>
[(1, DATE('2021-01-01'), 0),
(1, DATE('2021-01-05'), 5),
(1, DATE('2021-01-10'), 0),
(2, DATE('2021-01-03'), 10),
(2, DATE('2021-01-08'), 0),
(2, DATE('2021-01-09'), 0)])
advertisement campaign attribition data example:
SELECT
*
FROM UNNEST(ARRAY<STRUCT<ID INT64, Attribution_Timestamp Timestamp, campaign_name STRING>>
[(1, TIMESTAMP('2021-01-01 09:54:31'), "A"),
(1, TIMESTAMP('2021-01-09 22:32:51'), "B"),
(2, TIMESTAMP('2021-01-03 19:12:11'), "A")])
The end result I would like to get is:
SELECT
*
FROM UNNEST(ARRAY<STRUCT<ID INT64, UTC_Date DATE, Revenue FLOAT64, campaign_name STRING>>
[(1, DATE('2021-01-01'), 0, "A"),
(1, DATE('2021-01-05'), 5, "A"),
(1, DATE('2021-01-10'), 0, "B"),
(2, DATE('2021-01-03'), 10, "A"),
(2, DATE('2021-01-08'), 0, "A"),
(2, DATE('2021-01-09'), 0, "A")])
This can be achieved by somehow joining the campaign attribution data to the app data and then forward filling.
The problem I have is that the advertisment attribution timestamp can have a mismatch with the UTC dates in the app data table. This means I cannot use a left join as it will not assign campaign_name B to ID 1. Does anyone know an elegant way to solve this problem?

Answer 1

Found a solution! Here is what I did (and a little bit more sample data):
WITH app_data AS
(
SELECT
*
FROM UNNEST(ARRAY<STRUCT<adid INT64, utc_date DATE, Revenue FLOAT64>>
[(1, DATE('2021-01-01'), 0),
(1, DATE('2021-01-05'), 5),
(1, DATE('2021-01-10'), 0),
(1, DATE('2021-01-12'), 0),
(1, DATE('2021-01-15'), 0),
(1, DATE('2021-01-16'), 15),
(1, DATE('2021-01-18'), 0),
(2, DATE('2021-01-03'), 10),
(2, DATE('2021-01-08'), 0),
(2, DATE('2021-01-09'), 0),
(2, DATE('2021-01-15'), 4),
(2, DATE('2021-02-01'), 0),
(2, DATE('2021-02-08'), 8),
(2, DATE('2021-02-15'), 0),
(2, DATE('2021-03-04'), 0),
(2, DATE('2021-03-06'), 12),
(3, DATE('2021-02-15'), 10),
(3, DATE('2021-02-23'), 5),
(3, DATE('2021-03-25'), 0),
(3, DATE('2021-03-30'), 0)])
),
advertisment_attribution_data AS
(
SELECT
*
FROM UNNEST(ARRAY<STRUCT<adid INT64, utc_date DATE, campaign_name STRING>>
[(1, DATE(TIMESTAMP('2021-01-01 09:54:31')), "A"),
(1, DATE(TIMESTAMP('2021-01-09 22:32:51')), "B"),
(1, DATE(TIMESTAMP('2021-01-17 14:30:05')), "C"),
(2, DATE(TIMESTAMP('2021-01-03 19:12:11')), "A"),
(1, DATE(TIMESTAMP('2021-01-15 18:17:57')), "B"),
(3, DATE(TIMESTAMP('2021-03-14 22:32:51')), "C")])
)
SELECT
t1.*,
IFNULL(LAST_VALUE(t2.campaign_name IGNORE NULLS) OVER (PARTITION BY t1.adid ORDER BY t1.utc_date ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW), "Organic") as campaign_name
FROM
app_data t1
LEFT JOIN
advertisment_attribution_data t2
ON t1.adid = t2.adid
AND t1.utc_date = (SELECT MIN(t3.utc_date) FROM app_data t3 WHERE t2.adid=t3.adid AND t2.utc_date <= t3.utc_date)
EDIT
It doesn't work when I select a real table in app_data. It says: Unsupported subquery with table in join predicate.
EDIT 2
Found a way to solve the problem where you cannot use subqueries in joins (apparently it is possible for tables which are not selected from an existing table...) This is the way it works in any case:
WITH app_data AS
(
SELECT
*
FROM UNNEST(ARRAY<STRUCT<adid INT64, utc_date DATE, Revenue FLOAT64>>
[(1, DATE('2021-01-01'), 0),
(1, DATE('2021-01-05'), 5),
(1, DATE('2021-01-10'), 0),
(1, DATE('2021-01-12'), 0),
(1, DATE('2021-01-15'), 0),
(1, DATE('2021-01-16'), 15),
(1, DATE('2021-01-18'), 0),
(2, DATE('2021-01-03'), 10),
(2, DATE('2021-01-08'), 0),
(2, DATE('2021-01-09'), 0),
(2, DATE('2021-01-15'), 4),
(2, DATE('2021-02-01'), 0),
(2, DATE('2021-02-08'), 8),
(2, DATE('2021-02-15'), 0),
(2, DATE('2021-03-04'), 0),
(2, DATE('2021-03-06'), 12),
(3, DATE('2021-02-15'), 10),
(3, DATE('2021-02-23'), 5),
(3, DATE('2021-03-25'), 0),
(3, DATE('2021-03-30'), 0)])
),
advertisment_attribution_data AS
(
SELECT
*,
(
SELECT
MIN(t2.utc_date)
FROM app_data t2
WHERE t1.adid=t2.adid
AND t1.utc_date <= t2.utc_date
) as attribution_join_date -- is the closest next date for this adid in app_data to the attribution date. This ensures the join lateron works.
FROM UNNEST(ARRAY<STRUCT<adid INT64, utc_date DATE, campaign_name STRING>>
[(1, DATE(TIMESTAMP('2021-01-01 09:54:31')), "A"),
(1, DATE(TIMESTAMP('2021-01-09 22:32:51')), "B"),
(1, DATE(TIMESTAMP('2021-01-17 14:30:05')), "C"),
(2, DATE(TIMESTAMP('2021-01-03 19:12:11')), "A"),
(1, DATE(TIMESTAMP('2021-01-15 18:17:57')), "B"),
(3, DATE(TIMESTAMP('2021-03-14 22:32:51')), "C")]) t1
)
SELECT
t1.*,
IFNULL(LAST_VALUE(t2.campaign_name IGNORE NULLS) OVER (PARTITION BY t1.adid ORDER BY t1.utc_date ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW), 'Organic') as campaign_name
FROM
app_data t1
LEFT JOIN
advertisment_attribution_data t2
ON t1.adid = t2.adid
AND t1.utc_date = t2.attribution_join_date