QQ regarding autoscaling with adaptive scheduler (flink 1.18, k8s operator 1.8)
Started with initial parallelism of 64, after some time it scaled down to a maximum of 16.
Each TM is configured to have 16 CPU. But instead of releasing the nodes it still use all of them 4.
Anyways to improve that (e.g. to make autoscaler release at least one of them?)