Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
-
ghx-label-8
Description
We've seen several workloads recently where the preaggregations expanded to consume nearly all of the memory reservation on a host, starving the merge agg. We should think about runtime mechanisms to reduce the impact of this, e.g. stop expanding once we are low on memory.
I think we'd need to do this based on reservation_limit_ - child_reservations_ - used_reservation_ on the query mem tracker. Various heuristics are possible. Even something simple like keeping 10-20% free would be beneficial.
Related: IMPALA-2945 IMPALA-9530 IMPALA-9591
Attachments
Issue Links
- relates to
-
IMPALA-2945 Pre-aggregation cardinality estimates do not take into account data distribution
-
- Open
-
-
IMPALA-9591 Automatically cap pre-aggregation memory consumption to match estimate
-
- Open
-
-
IMPALA-9530 Allow limiting memory consumed by preaggregation
-
- Resolved
-