Details
-
New Feature
-
Status: Resolved
-
Major
-
Resolution: Won't Fix
-
Impala 2.5.0
-
None
Description
When running complex queries on large clusters with lots of runtime filters the coordinator quickly becomes network bound due to the extra incoming and outgoing traffic for runtime filters, once the coordinator becomes network bound all other fragments in the cluster are negatively affected as they get blocked on shuffling/broadcasting data to the coordinator node.
This bottleneck was identified when running large scale tests on EC2 nodes with less than ideal network throughput.
In attached png is aggregate network throughput across the 32 nodes in the cluster with the coordinator in red.
Compression should alleviate this bottleneck but we should consider other solutions
Attachments
Attachments
Issue Links
- is related to
-
IMPALA-3825 Distribute runtime filter aggregation across cluster
- Resolved
-
IMPALA-3610 Track non-RPC memory from global runtime filters on the coordinator
- Resolved