Pod Priority and Preemption

When a cluster fills up, which pod wins the last free CPU/memory — your Prometheus monitoring stack or a nightly batch job? Without priority, scheduling is essentially first-come. Pod priority ranks pending pods, and preemption lets a high-priority pod (here, Prometheus) evict lower-priority ones when there is otherwise no room.

Solution: Define PriorityClasses, stamp pods with a priorityClassName, and let the scheduler order the pending queue by priority and preempt lower-priority victims when a high-priority pod cannot otherwise be placed.

When the last berth at a full port is needed by a VIP cargo ship, the harbor master ranks the waiting ships and bumps a lower-priority vessel off the berth so the VIP can dock, the displaced ship must wait or move on. Lower-ranked cargo never displaces equal or higher-ranked cargo. Kubernetes scheduling mirrors this: a PriorityClass ranks pods, the scheduler orders the pending queue by priority, and when a high-priority pod (Prometheus) cannot fit on a full node it preempts, evicting the lowest-priority victim to make room.

Graph legend — each Kubernetes node maps a port concept to the real Prometheus preemption case:

The scheduler does not place pods at random — it runs each pending pod through a short pipeline (queue, filter, score, pick), and priority is what decides a pod's place in line. The diagram below shows where priority and preemption (evicting lower-ranked pods to make room) slot into that cycle: a high-priority pod like Prometheus jumps to the front of the queue and, if no node fits, triggers the preemption step.

Graph legend — each node is a stage of the scheduling cycle for the Prometheus pod:

A PriorityClass is cluster-scoped. Higher value = higher priority. Values above one billion are reserved for system-critical classes (system-cluster-critical, system-node-critical).

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: high-priority
value: 1000000
globalDefault: false
description: "Critical platform workloads (Prometheus, Alertmanager)"
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: low-priority
value: 1000
globalDefault: false
preemptionPolicy: Never        # this pod waits, never evicts others
description: "Best-effort batch jobs"

kubectl get priorityclass                # built-ins + yours
kubectl describe priorityclass high-priority

Reference the class by name; the admission controller stamps the numeric priority onto the pod spec. Here a real Prometheus server claims high priority.

apiVersion: v1
kind: Pod
metadata:
  name: prometheus
  labels: { app: prometheus }
spec:
  priorityClassName: high-priority
  containers:
    - name: prometheus
      image: prom/prometheus:v2.54.0
      args: ["--config.file=/etc/prometheus/prometheus.yml"]
      ports:
        - { containerPort: 9090, name: web }
      resources:
        requests:
          cpu: "1"
          memory: 1Gi

kubectl get pod prometheus -o jsonpath='{.spec.priority}{"\n"}'   # 1000000

Two fields decide how a class behaves when pods omit a priority and whether it may evict others. Here is what each one does:

   globalDefault: true   -> pods WITHOUT a priorityClassName get this value
                            (only ONE class should set this; default otherwise = 0)
   preemptionPolicy:
     PreemptLowerPriority (default) -> may evict lower-priority pods to fit
     Never                          -> scheduled by priority order, but NEVER evicts;
                                       useful for high-priority but non-urgent jobs

When the Prometheus pod is Pending and no node fits, the scheduler looks for a node where evicting one or more lower-priority pods would let it fit, picking the set with the least disruption. Victims get a graceful termination.

Graph legend — each node is a real element of the victim-selection step:

kubectl get events --field-selector reason=Preempted
kubectl describe pod prometheus | grep -i preempt

This diagram traces the full life of one high-priority pod, from the moment you create it to the moment it lands on a node. The key branch is what happens when no node has room: depending on the pod's preemptionPolicy (whether it is allowed to evict others), it either waits or forces lower-priority pods off a node to take their place.

Graph legend — each node is a real step from pod creation to scheduling:

A complete walkthrough: define two PriorityClasses, saturate a node with low-priority batch pods, then schedule the high-priority Prometheus server that cannot fit and watch the scheduler evict a low-priority victim to make room.

1. Create the two PriorityClasses:

apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: low-priority
value: 1000
globalDefault: false
description: "Best-effort batch jobs"
---
apiVersion: scheduling.k8s.io/v1
kind: PriorityClass
metadata:
  name: high-priority
value: 1000000
globalDefault: false
description: "Critical platform workloads (Prometheus)"

kubectl get priorityclass low-priority high-priority
# expected: both listed with values 1000 and 1000000

2. Pick one node and size requests so three low-priority batch pods nearly fill its allocatable CPU:

kubectl get nodes
NODE=$(kubectl get nodes -o jsonpath='{.items[0].metadata.name}')
kubectl describe node "$NODE" | grep -A4 Allocatable   # note cpu, e.g. 4

# low-fill.yaml  -> 3 replicas, each requesting ~1 CPU, pinned to the chosen node
apiVersion: apps/v1
kind: Deployment
metadata:
  name: batch-filler
spec:
  replicas: 3
  selector:
    matchLabels: { app: batch-filler }
  template:
    metadata:
      labels: { app: batch-filler }
    spec:
      priorityClassName: low-priority
      nodeName: NODE_PLACEHOLDER
      containers:
        - name: pause
          image: registry.k8s.io/pause:3.9
          resources:
            requests:
              cpu: "1"

3. Substitute the node name and apply; confirm the node is now nearly full:

sed "s/NODE_PLACEHOLDER/$NODE/" low-fill.yaml | kubectl apply -f -
kubectl get pods -l app=batch-filler -o wide
# expected: 3 batch-filler pods Running on $NODE

kubectl describe node "$NODE" | grep -A3 "Allocated resources"
# expected: cpu requests near 100% of allocatable

4. Schedule the Prometheus pod with high priority and a request that cannot fit without eviction:

apiVersion: v1
kind: Pod
metadata:
  name: prometheus
  labels: { app: prometheus }
spec:
  priorityClassName: high-priority
  nodeName: NODE_PLACEHOLDER
  containers:
    - name: prometheus
      image: prom/prometheus:v2.54.0
      args: ["--config.file=/etc/prometheus/prometheus.yml"]
      resources:
        requests:
          cpu: "1"

sed "s/NODE_PLACEHOLDER/$NODE/" prometheus.yaml | kubectl apply -f -

5. Watch the scheduler preempt a low-priority victim, then bind Prometheus:

kubectl get events --field-selector reason=Preempted
# expected: pod/batch-filler-xxxx Preempted by default/prometheus on node $NODE

kubectl get pods -l app=batch-filler
# expected: one batch-filler pod gone (or Pending, recreated by the Deployment)

kubectl get pod prometheus -o wide
# expected: prometheus Running on $NODE

6. Inspect the priority that was stamped and confirm equal/higher pods are never victims:

kubectl get pod prometheus -o jsonpath='{.spec.priority}{"\n"}'   # 1000000
kubectl describe pod prometheus | grep -i preempt
# expected: nominated node + preemption note for the victim

Graph legend — each node is a real step in the fill-then-preempt walkthrough:

These are the priority and preemption mistakes that bite most often:

   - priorityClassName typo        -> pod rejected; the class must already exist
   - two globalDefault classes     -> only one may set globalDefault: true
   - surprise evictions            -> a high-priority pod silently preempted batch jobs
   - editing priority on a pod     -> immutable after creation; recreate the pod
   - PDB ignored under pressure    -> PDBs are best-effort during preemption
   - values >1e9                   -> reserved for system-critical classes

• A PriorityClass maps a name to an integer; higher = scheduled first.
• Pods reference it via priorityClassName; the value is immutable.
• globalDefault sets the priority for pods that name no class (else 0).
• Preemption evicts the fewest, lowest-priority victims to fit a higher-priority pod.
• preemptionPolicy: Never lets a pod jump the queue but never evict others.
• Values above one billion are reserved for system-critical workloads.

• [ ] Created high- and low-priority PriorityClasses
• [ ] Assigned priorityClassName to the Prometheus pod and read .spec.priority
• [ ] Filled a node and watched Prometheus preempt a batch victim
• [ ] Set preemptionPolicy: Never and confirmed it never evicts
• [ ] Explained globalDefault and the system-critical reserved range