Open zwForrest opened 3 weeks ago
Nice proposal! koord-descheduler re-uses the NodeFit pre-check logic of Kubernetes Descheduler's DefaultEvictor. This is not enough in all real cases. Other optimization to NodeFit includes like TopologySpreadConstraints and so on. We'll continue to improve this.
What is your proposal: Currently, during eviction in rescheduling, only the total cluster resources are considered to determine if they are sufficient for the pod being rescheduled, without taking into account whether single nodes have the required resources.
Why is this needed: Prevent the scenario where other nodes become hotspots after rescheduling.
Is there a suggested solution, if so, please add it: