Postgresql: Grouping with limit on group size using window functions - postgresql

Question

Is there a way in Postgresql to write a query which groups rows based on a column with a limit without discarding additional rows.
Say I've got a table with three columns id, color, score with the following rows
1 red 10.0
2 red 7.0
3 red 3.0
4 blue 5.0
5 green 4.0
6 blue 2.0
7 blue 1.0
I can get a grouping based on color with window functions with the following query
SELECT * FROM (
SELECT id, color, score, rank()
OVER (PARTITION BY color ORDER BY score DESC)
FROM grouping_test
) AS foo WHERE rank <= 2;
with the result
id | color | score | rank
----+-------+-------+------
4 | blue | 5.0 | 1
6 | blue | 2.0 | 2
5 | green | 4.0 | 1
1 | red | 10.0 | 1
2 | red | 7.0 | 2
which discards item with ranks > 2. However what I need is a result like
1 red 10.0
2 red 7.0
4 blue 5.0
6 blue 2.0
5 green 4.0
3 red 3.0
7 blue 1.0
With no discarded rows.
Edit:
To be more precise about the logic I need:
Get me the row with the highest score
The next row with the same color and the highest possible score
The item with the highest score of the remaining items
Same as 2., but for the row from 3.
...
Continue as long as pairs with the same color can be found, then order whats left by descending score.
The import statements for a test table can be found here.
Thanks for your help.

Answer 1

It can be done using two nested window functions
SELECT
id
FROM (
SELECT
id,
color,
score,
((rank() OVER color_window) - 1) / 2 AS rank_window_id
FROM grouping_test
WINDOW color_window AS (PARTITION BY color ORDER BY score DESC)
) as foo
WINDOW rank_window AS (PARTITION BY (color, rank_window_id))
ORDER BY
(max(score) OVER rank_window) DESC,
color;
With 2 being the parameter of the group size.

Answer 2

You can do ORDER BY (rank <= 2) DESC to get the rows with rank<=2 above all else:
SELECT id,color,score FROM (
SELECT id, color, score, rank()
OVER (PARTITION BY color ORDER BY score DESC),
max(score) OVER (PARTITION BY color) mx
FROM grouping_test
) AS foo
ORDER BY
(rank <= 2) DESC,
CASE WHEN rank<=2 THEN mx ELSE NULL END DESC,
id;
http://sqlfiddle.com/#!12/bbcfc/109