Select Last Rows with Distinct Field

I have a table with the following schema:

id itemid date        some additional data
1   1000  10/12/2020       a
2   1000  10/12/2020       b
3   1002  09/12/2020       c
4   1001  07/12/2020       d
5   1000  05/12/2020       e
6   1005  03/12/2020       f
7   1003  03/12/2020       g

In this table only the id field is unique. I’m concerned with getting the rows containing the last X distinct itemid, ordered by date.

For example, in the sample above, if I’d like to get the last 3 distinct itemid, I’d be getting the first 4 rows, since in the first 4 rows we have three distinct itemid: 1000, 1002 and 1001. I’m not sure how to achieve this using a single SQL statement.

Answer

If I understand correctly, you would like to count the number of distinct item ids up to each each row (by date) and return all rows where the count is three.

If Postgres supported this, you could use:

select t.*
from (select t.*, 
             count(*) filter (where id = min_id) over (order by date desc) as cnt_itemid
      from (select t.*,
                   min(id) over (partition by itemid order by date desc) as min_id
            from t
           ) t
     ) t
where cnt_itemid <= 3;

Alas, Postgres does not support COUNT(DISTINCT) as a window function. But you can calculate it using DENSE_RANK():

select t.*
from (select t.*, 
             count(*) over (filter where id = min_id) as cnt_itemid
      from (select t.*,
                   min(id) over (partition by itemid order by date) as min_id
            from t
           ) t
     ) t
where cnt_itemid <= 3;

However, this returns all the most recent rows up before the 4th item — so it has extra rows.

To get four rows, you want the first where the item id is “3”. One method is:

select t.*
from (select t.*, min(id) filter (where cnt_itemid = 3) over () as min_cnt_itemid_3
      from (select t.*, 
                   count(*) filter (where id = min_id) over (order by date desc) as cnt_itemid
            from (select t.*,
                         min(id) over (partition by itemid order by date desc) as min_id
                  from t
                 ) t
           ) t
     ) t
where id <= min_cnt_itemid_3;

You can also do this by identifying the first occurrence of the “third item” and then choosing all rows up to that row:

select t.*
from t join
     (select itemid, min(max_date) over () as min_max_date
      from (select t.itemid, max(date) as max_date
            from t
            group by t.itemid
            order by max(t.date) desc
            limit 3
           ) t
      ) tt
      on t.itemid = tt.itemid and t.date >= tt.min_max_date;

This fiddle shows each of these.