Priority

As shown in process, part of the scheduling depends on the defined priority of the pods to be scheduled. To control this, k8s allows for the definition of objects called PriorityClass which determined how critical a pod is.

The range of priorities that can be defined, can go from 10^9 to -2^31 (-2,147,483,648). Priority classes are also used for system components deployed (thing of the kube-system namespace), and these ones go from 10^9 to 2*10^9 (and maybe more).

Once defined, priority classes are global to the cluster (e.g not attached to namespaces). By default, the two priorities used for system components exist (system-cluster-critical, system-node-critical)

Very importantly, the classes allow to define how to behave with respect to lower priority pods. The options are: Preempt them (PreemptLowerPriority) or not (never) with the default being Preemption.

Tip

By default, all pods that do not defined a priority class, have priority 0, this can be changed assigning a priority class as default with globalDefault

Definition

To define it, the following YAML structure is used

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
    name: my-prio
value: 1000000000
description: "Ultra high prio tasks"
preemptionPolicy: PreemptLowerPriority
globalDefault:

Once defined, pods can be created referencing a class:

apiVersion: v1
kind: Pod
metadata:
    name: mypod
spec:
    priorityClassName: my-prio