When we do high concurrency test to Doris, we found that one of all backends are complete consumed, but other backends have 60% CPU idle.

The test query likes "select day, sum(pv) from tbl where day = 20190904 group by day". After looking into backend's log, I found that all queries' second phase aggregation was processed in the exhausted backend.

This is because Doris will assign the same backend for aggregation/join shuffle slot, this will cause compute skew when the query has smaller group by cardinality than number of backends.