Gyula Fora created FLINK-30466: ---------------------------------- Summary: Detect ineffective scaling operations Key: FLINK-30466 URL: https://issues.apache.org/jira/browse/FLINK-30466 Project: Flink Issue Type: New Feature Components: Kubernetes Operator Reporter: Gyula Fora Assignee: Matyas Orhidi Fix For: kubernetes-operator-1.4.0
The autoscaler logic currently does not consider whether a given proposed parallelism change will achieve the desired processing rate based on historic metrics. It might be possible that we are for example trying to scale down to a parallelism that we have also done in the past which resulted in an immediate scale up. This could be avoided by comparing the expected processing rate at a certain parallelism with past metrics from that parallelism setting. There should be some flexibility and a limited "memory" but some simple checks could reduce instable scaling caused by non-linear performance changes. -- This message was sent by Atlassian Jira (v8.20.10#820010)