Running sum of unique users in redshift

I have a table with as follows with user visits by day -

| date     | user_id  |
|:-------- |:-------- |
| 01/31/23 | a        |
| 01/31/23 | a        |
| 01/31/23 | b        |
| 01/30/23 | c        |
| 01/30/23 | a        |
| 01/29/23 | c        |
| 01/28/23 | d        |
| 01/28/23 | e        |
| 01/01/23 | a        |
| 12/31/22 | c        |

I am looking to get a running total of unique user_id for the last 30 days . Here is the expected output -

| date     | distinct_users|
|:-------- |:--------      |
| 01/31/23 | 5             |
| 01/30/23 | 4             |
.
.
.

Here is the query I tried -

SELECT date
     , SUM(COUNT(DISTINCT user_id)) over (order by date rows between 30 preceding and current row) AS unique_users
FROM mytable
GROUP BY date
ORDER BY date DESC

The problem I am running into is that this query not counting the unique user_id - for instance the result I am getting for 01/31/23 is 9 instead of 5 as it is counting user_id 'a' every time it occurs.

Thank you, appreciate your help!

Not the most performant approach, but you could use a correlated subquery to find the distinct count of users over a window of the past 30 days:

<!-- language: sql -->

SELECT
    date,
    (SELECT COUNT(DISTINCT t2.user_id)
     FROM mytable t2
     WHERE t2.date BETWEEN t1.date - INTERVAL '30 day' AND t1.date) AS distinct_users
FROM mytable t1
ORDER BY date;

There are a few things going on here. First window functions run after group by and aggregation. So COUNT(DISTINCT user_id) gives the count of user_ids for each date then the window function runs. Also, window function set up like this work over the past 30 rows, not 30 days so you will need to fill in missing dates to use them.

As to how to do this - I can only think of the "expand to the data so each date and id has a row" method. This will require a CTE to generate the last 2 years of dates plus 30 days so that the look-back window works for the first dates. Then window over the past 30 days for each user_id and date to see which rows have an example of this user_id within the past 30 days, setting the value to NULL if no uses of the user_id are present within the window. Then Count the user_ids counts (non NULL) grouping by just date to get the number of unique user_ids for that date.

This means expanding the data significantly but I see no other way to get truly unique user_ids over the past 30 days. I can help code this up if you need but will look something like:

WITH RECURSIVE CTE to generate the needed dates,
CTE to cross join these dates with a distinct set of all the user_ids in user for the past 2 years,
CTE to join the date/user_id data set with the table of real data for past 2 years and 30 days and window back counting non-NULL user_ids, partition by date and user_id, order by date, and setting any zero counts to NULL with a DECODE() or CASE statement,
SELECT, grouping by just date count the user_ids by date;