Split out count distincts and other distinct aggs too - when mixed with other aggs - into a seperate step and join

[kaikalur]

Sometimes count distinct can make the plan too complex. So it will be good to split out the count distinct into a separate subplan and join. Roughly in SQL:

SELECT COUNT(DISTINCT x) v1, COUNT(DISTINCT y) AS v2, approx_set(x) as s1, Approx_set(y) s2 FROM T group by k

can be done as

SELECT * FROM
(SELECT k, COUNT(DISTINCT x) v1, COUNT(DISTINCT y) AS v2 FROM T group by k)
JOIN
(SELECT k, approx_set(x) as s1, Approx_set(y) s2 FROM T group by k)
USING(k)

(Need to take care of nulls - maybe by coalescing or something)

[kaikalur]

Or one can do full join/union and do another arbitrary(v) v on all the values for more robustness and ease of handling nulls

SELECT k, arbitrary(v1) v1, arbitrary(v2) v2, arbitrary(s1) s1, arbitrary(s2) s2, FROM
(
SELECT k, COUNT(DISTINCT x) v1, COUNT(DISTINCT y) AS v2, null s1, null s2 FROM T group by k
union all
SELECT k, null v1, null v2, approx_set(x) as s1, Approx_set(y) s2 FROM T group by k
)
group by k

[kaikalur]

Or even separate each distinct into its own subselect and add - depending on the number of distinct values being agg