Requests decide whether a Pod fits. Priority decides who gives way when it does not.

A PriorityClass

kubectl create priorityclass high --value=1000 --description="urgent work"
kubectl get priorityclass

NAME                      VALUE        GLOBAL-DEFAULT   AGE
high                      1000         false            5s
system-cluster-critical   2000000000   false            12d
system-node-critical      2000001000   false            12d

A PriorityClass is cluster-scoped and holds a number. The two system classes exist on every cluster; control-plane Pods use them so nothing you create can push them out.

Using it

spec:
  priorityClassName: high
  containers:
    - name: job
      image: stackmonsters/monster-market-worker

At admission the API server looks the class up and writes spec.priority: 1000 into the Pod. Pods without a class get 0, unless one class is marked globalDefault. A name that does not exist is an error: no PriorityClass with name … was found.

Preemption

When a Pod cannot be scheduled and its priority is higher than some Pods already running, the scheduler looks for a node where evicting lower-priority Pods would free enough room. It picks the node where the fewest have to go, deletes them, and places the new Pod there once they are gone.

The victims get an event:

Normal  Preempted  Preempted by peaks/urgent on node peaks-node-2

If a controller owns them, it recreates them, and they queue behind the urgent Pod. A bare Pod is simply gone.

Reading it

kubectl describe pod urgent | grep -i priority
kubectl get events --field-selector reason=Preempted

Priority:             1000
Priority Class Name:  high