What would you like to be added:

Support Scheduling Priority and Preemption

Why is this needed:

To further improve the packing rate and resource utilization of the cluster, we usually divide tasks into high-priority tasks and low-priority tasks, and deploy them in a mixed manner. When the cluster is idle, low-priority tasks can make full use of cluster resources, while high-priority tasks can preempt low-priority task resources to meet the needs of high-priority tasks. In a single Kubernetes cluster, the community provides the pod priority feature to implement the scheduling priority of pods, and also implements preemptive scheduling of pods in the kube-scheduler based on priority.

When we move the application deployment process to a multi-cluster level, such scenarios still exist. However, in the current Clusternet multi-cluster application management mechanism, no similar capabilities are provided.

The entire requirement can be divided into two parts: "scheduling priority" and "preemptive scheduling".

"Scheduling Priority" implementation:

Add a scheduling priority field in the subscription
Add a priority queue in the scheduler. After the scheduler listens to the subscription changes, it puts them into the priority queue and dequeues them according to priority.

"preemptive scheduling"

There are two implementation ideas for "preemptive scheduling":

Idea 1: Implement subscription preemption;

• step1: Reuse the "scheduling priority" implementation;
• step2: Extend the preemption logic based on the existing clusternet-scheduler provided scheduler framework;

Idea 2: Rely on single-cluster preemption logic;

• step1: Enhance the clusternet-scheduler's resource prediction process's perception of pod priority, It is necessary to supplement pod priority as scheduling information in feedinventory;
• step2: Enhance the cross-cluster rescheduling capability of low-priority pods;

btw: I prefer idea 1