Mastering Kubernetes QoS: Guarantees, Burstable, and BestEffort Explained

Kubernetes QoS Overview

QoS (Quality of Service) classes in Kubernetes determine pod scheduling and eviction priority. Kubelet uses them to manage pod eviction order and advanced CPU management.

QoS Levels

Guaranteed : All containers in the pod must have identical memory and CPU limits and requests.

Burstable : At least one container has a memory or CPU request.

BestEffort : No container specifies any memory or CPU requests or limits.

Guaranteed

For a Guaranteed pod, each container (including init containers) must specify equal memory request and limit, and equal CPU request and limit.

apiVersion: v1
kind: Pod
metadata:
  name: qos-demo
spec:
  containers:
  - name: qos-demo
    image: nginx
    resources:
      limits:
        memory: "500Mi"
        cpu: "700m"
      requests:
        memory: "500Mi"
        cpu: "700m"

Verification:

# kubectl describe po qos-demo
...
QoS Class:               Guaranteed

If a container sets limits without requests, Kubernetes automatically sets matching requests.

Burstable

A pod is classified as Burstable when it does not meet Guaranteed criteria and at least one container has a memory or CPU request.

apiVersion: v1
kind: Pod
metadata:
  name: qos-demo2
spec:
  containers:
  - name: qos-demo2
    image: nginx
    resources:
      limits:
        memory: "500Mi"
      requests:
        memory: "200Mi"

Verification:

# kubectl describe po qos-demo2
...
QoS Class:               Burstable

BestEffort

For BestEffort pods, containers have no memory or CPU limits or requests.

apiVersion: v1
kind: Pod
metadata:
  name: qos-demo3
spec:
  containers:
  - name: qos-demo3
    image: nginx

QoS Priority

The priority order from lowest to highest is: BestEffort → Burstable → Guaranteed.

Eviction Mechanism

Compressible resource: CPU . When a pod exceeds its CPU limit, usage is throttled but not killed.

Non‑compressible resource: Memory . If a node runs out of memory before Kubelet can reclaim it, the OOM killer terminates containers based on their oom_score_adj values: -998 for Guaranteed, 1000 for BestEffort, and a calculated value for Burstable.

The OOM killer first terminates the lowest‑QoS pods that exceed their requested resources, typically selecting over‑requesting containers in Burstable pods.

Best Practices

Classify applications by type: core, regular, extra.

Assign QoS classes: core → Guaranteed, regular → Burstable, extra → BestEffort.

Separate cluster nodes accordingly: core nodes, regular nodes, extra nodes.

Scheduling policies:

Core apps: use nodeAffinity with preferredDuringSchedulingIgnoredDuringExecution to target core nodes.

Regular apps: use hard nodeAffinity to schedule on regular nodes.

Extra apps: use hard nodeAffinity to schedule on extra nodes.