Rolling 90 days active users in BigQuery, improving preformance (DAU/MAU/WAU)

I'm trying to get the number of unique events on a specific date, rolling 90/30/7 days back. I've got this working on a limited number of rows with the query bellow but for large data sets I get memory errors from the aggregated string which becomes massive.

I'm looking for a more effective way of achieving the same result.

Table looks something like this:

+---+------------+-------------+
|   |     date   |     userid  |
+---+------------+-------------+
| 1 | 2013-05-14 | xxxxx       |
| 2 | 2017-03-14 | xxxxx       |
| 3 | 2018-01-24 | xxxxx       |
| 4 | 2013-03-21 | xxxxx       |
| 5 | 2014-03-19 | xxxxx       |
| 6 | 2015-09-03 | xxxxx       |
| 7 | 2014-02-06 | xxxxx       |
| 8 | 2014-10-30 | xxxxx       |
| ..| ...        | ...         |
+---+------------+-------------+

Format of the desired result:

+---+------------+---------------------------------------------+
|   |     date   | active_users_7_days | active_users_90_days  |
+---+------------+---------------------------------------------+
| 1 | 2013-05-14 | 1240                | 34339                 |
| 2 | 2017-03-14 | 4334                | 54343                 |
| 3 | 2018-01-24 | .....               | .....                 |
| 4 | 2013-03-21 | .....               | .....                 |
| 5 | 2014-03-19 | .....               | .....                 |
| 6 | 2015-09-03 | .....               | .....                 |
| 7 | 2014-02-06 | .....               | .....                 |
| 8 | 2014-10-30 | .....               | .....                 |
| ..| ...        | .....               | .....                 |
+---+------------+---------------------------------------------+

My query looks like this:

#standardSQL
    WITH
      T1 AS(
      SELECT
        date,
        STRING_AGG(DISTINCT userid) AS IDs
      FROM
        `consumer.events`
      GROUP BY
        date ),
      T2 AS(
      SELECT
        date,
        STRING_AGG(IDs) OVER(ORDER BY UNIX_DATE(date) RANGE BETWEEN 90 PRECEDING
          AND CURRENT ROW) AS IDs
      FROM
        T1 )
    SELECT
      date,
      (
      SELECT
        COUNT(DISTINCT (userid))
      FROM
        UNNEST(SPLIT(IDs)) AS userid) AS NinetyDays
    FROM
      T2

Solution 1:

Counting unique users requires a lot of resources, even more if you want results over a rolling window. For a scalable solution, look into approximate algorithms like HLL++:

• https://medium.freecodecamp.org/counting-uniques-faster-in-bigquery-with-hyperloglog-5d3764493a5a

For an exact count, this would work (but gets slower as the window gets larger):

#standardSQL
SELECT DATE_SUB(date, INTERVAL i DAY) date_grp
 , COUNT(DISTINCT owner_user_id) unique_90_day_users
 , COUNT(DISTINCT IF(i<31,owner_user_id,null)) unique_30_day_users
 , COUNT(DISTINCT IF(i<8,owner_user_id,null)) unique_7_day_users
FROM (
  SELECT DATE(creation_date) date, owner_user_id
  FROM `bigquery-public-data.stackoverflow.posts_questions` 
  WHERE EXTRACT(YEAR FROM creation_date)=2017
  GROUP BY 1, 2
), UNNEST(GENERATE_ARRAY(1, 90)) i
GROUP BY 1
ORDER BY date_grp

The approximate solution produces results way faster (14s vs 366s, but then the results are approximate):

#standardSQL
SELECT DATE_SUB(date, INTERVAL i DAY) date_grp
 , HLL_COUNT.MERGE(sketch) unique_90_day_users
 , HLL_COUNT.MERGE(DISTINCT IF(i<31,sketch,null)) unique_30_day_users
 , HLL_COUNT.MERGE(DISTINCT IF(i<8,sketch,null)) unique_7_day_users
FROM (
  SELECT DATE(creation_date) date, HLL_COUNT.INIT(owner_user_id) sketch
  FROM `bigquery-public-data.stackoverflow.posts_questions` 
  WHERE EXTRACT(YEAR FROM creation_date)=2017
  GROUP BY 1
), UNNEST(GENERATE_ARRAY(1, 90)) i
GROUP BY 1
ORDER BY date_grp

Updated query that gives correct results - removing rows with less than 90 days (works when no dates are missing):

#standardSQL
SELECT DATE_SUB(date, INTERVAL i DAY) date_grp
 , HLL_COUNT.MERGE(sketch) unique_90_day_users
 , HLL_COUNT.MERGE(DISTINCT IF(i<31,sketch,null)) unique_30_day_users
 , HLL_COUNT.MERGE(DISTINCT IF(i<8,sketch,null)) unique_7_day_users
 , COUNT(*) window_days
FROM (
  SELECT DATE(creation_date) date, HLL_COUNT.INIT(owner_user_id) sketch
  FROM `bigquery-public-data.stackoverflow.posts_questions` 
  WHERE EXTRACT(YEAR FROM creation_date)=2017
  GROUP BY 1
), UNNEST(GENERATE_ARRAY(1, 90)) i
GROUP BY 1
HAVING window_days=90
ORDER BY date_grp

Solution 2:

You can aggregate the date and do the sum. What is the aggregation? Take the most recent date:

select count(*) as num_users,
       sum(case when date > datediff(current_date, interval -30 day) then 1 else 0 end) as num_users_30days,
       sum(case when date > datediff(current_date, interval -60 day) then 1 else 0 end) as num_users_60days,
       sum(case when date > datediff(current_date, interval -90 day) then 1 else 0 end) as num_users_90days
from (select user_id, max(date) as max(date)
      from `consumer.events` e
      group by user_id
     ) e;

If the most recent date for the user is in the period, then the user should be counted.

You can get this "as-of" a particular date by using a where clause in the subquery.