Wrong results: pushing a volatile filter below GROUP BY changes the count #27074

A filter applied after GROUP BY can change the number of rows counted within a group. Reproduced with DataFusion CLI 57.0.0-dev:

SELECT c
FROM (
    SELECT a, COUNT(*) AS c
    FROM (VALUES (1), (1)) AS t(a)
    GROUP BY a
) AS counts
WHERE random() < 0.5;

Expected: Both input rows belong to the same group. The query should return either 2 or no rows, depending on whether the group passes the filter.

Actual: The query can return 1. EXPLAIN VERBOSE shows that push_down_filter moves random() < 0.5 below the aggregate, so it runs once per input row. If exactly one row passes, the count is 1.

@DKER2 originally reported this behavior for an aggregate without GROUP BY. The example above reproduces it with an ordinary grouping key, independently of the empty-input behavior tracked in #26545. Original comment