You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
When doing more than one distinct aggregation Spark will currently insert in an ExpandExec followed by two Aggregation passes. This is to let us do the distinct aggregations at the same time as the non-distinct aggregations. It gets a little complicated. In some cases our HashAggregate actually ends up being a sort aggregation inside of CUDF because hash aggregations only work on a small set of app/type combinations. It would be really great if we could have a way for HashAggregate to indicate that the aggregations that are going to be done would result in a sort based aggregation. Then from that we could have upstream operators, like GpuExpandExec, recognize this and optionally sort the input batches (not the full input data) so that it satisfies the desired ordering. Then we could have a way for it to signal to GpuHashAggregate that the data is sorted by batches, which would let it avoid doing the sort in CUDF all together.
The text was updated successfully, but these errors were encountered:
Is your feature request related to a problem? Please describe.
When doing more than one distinct aggregation Spark will currently insert in an ExpandExec followed by two Aggregation passes. This is to let us do the distinct aggregations at the same time as the non-distinct aggregations. It gets a little complicated. In some cases our HashAggregate actually ends up being a sort aggregation inside of CUDF because hash aggregations only work on a small set of app/type combinations. It would be really great if we could have a way for HashAggregate to indicate that the aggregations that are going to be done would result in a sort based aggregation. Then from that we could have upstream operators, like GpuExpandExec, recognize this and optionally sort the input batches (not the full input data) so that it satisfies the desired ordering. Then we could have a way for it to signal to GpuHashAggregate that the data is sorted by batches, which would let it avoid doing the sort in CUDF all together.
The text was updated successfully, but these errors were encountered: