Kubernetes QoS Classes: Guaranteed, Burstable, and BestEffort Pod Eviction Priorities

QoS Class Assignment (automatic — you don't set it directly):

Guaranteed ← BEST: Last to be evicted
  Condition: requests == limits for ALL containers (both CPU and memory)
  resources:
    requests: {cpu: "500m", memory: "512Mi"}
    limits:   {cpu: "500m", memory: "512Mi"}  ← exactly equal

Burstable ← MIDDLE: Evicted after BestEffort
  Condition: At least one container has requests < limits
  OR: requests set but no limits (common pattern)
  resources:
    requests: {cpu: "250m", memory: "256Mi"}
    limits:   {memory: "512Mi"}               ← different from requests

BestEffort ← WORST: Evicted FIRST
  Condition: No requests OR limits set on ANY container
  resources: {}                               ← completely empty

# Check QoS class of a pod
kubectl get pod my-pod \
  -o jsonpath='{.status.qosClass}'
# Returns: Guaranteed | Burstable | BestEffort

# Check all pods and their QoS
kubectl get pods -A \
  -o custom-columns='NAMESPACE:.metadata.namespace,NAME:.metadata.name,QOS:.status.qosClass'

Node running low on memory:
        │
        ▼
kubelet evicts BestEffort pods first (highest consumption over request)
        │
        ▼
If still low: evicts Burstable pods (furthest over their request)
        │
        ▼
If still critical: evicts Guaranteed pods (last resort)
        │
        ▼
If node still unstable: kernel OOM killer starts killing processes

# Guaranteed QoS — for critical stateful workloads
resources:
  requests:
    cpu: "1000m"
    memory: "2Gi"
  limits:
    cpu: "1000m"      # Must equal requests for Guaranteed
    memory: "2Gi"     # Must equal requests for Guaranteed

# Burstable QoS — for most production stateless apps
resources:
  requests:
    cpu: "250m"
    memory: "256Mi"
  limits:
    memory: "512Mi"   # Higher than request = Burstable
    # No CPU limit = also Burstable (requests set but not limits)