Details
-
Improvement
-
Status: Resolved
-
Minor
-
Resolution: Fixed
-
Impala 4.2.0
-
None
-
ghx-label-3
Description
Currently, when updating an executor group's min size, we need to restart all executors in this executor group and all coordinators. Restarting coordinators will interrupt queries running on all executor groups.
It is necessary to restart coordinators because executor groups will never be removed from a cluster during the lifetime of an impalad and inconsistent executor group configurations will be ignored when adding a new executor: https://github.com/apache/impala/blob/c3ed44268838f4a00624e7e7fbf6e614f0a54a48/be/src/scheduling/executor-group.cc#L89-L93
We can improve this by removing empty/unused executor groups from the membership, and then new configurations of the same group names can be applied.