Autoscaling & scheduling: HPA, affinity, taints & tolerations

kubectl scale works, but someone has to be watching and typing it. This lesson covers letting Kubernetes make that call itself — and steering where pods land once it does.

Horizontal Pod Autoscaler (HPA)

An HPA watches a metric (CPU usage by default) and adjusts a Deployment's replica count automatically to keep it near a target.

# the quick way
kubectl autoscale deployment web --min=2 --max=10 --cpu-percent=70

# the declarative way — hpa.yaml
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
  name: web
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: web
  minReplicas: 2
  maxReplicas: 10
  metrics:
    - type: Resource
      resource:
        name: cpu
        target: { type: Utilization, averageUtilization: 70 }

Watch it in action with kubectl get hpa — under load, the replica count climbs on its own; once load drops, it scales back down.

Node affinity — pods that prefer (or require) certain nodes

Some pods have real reasons to prefer specific nodes — a GPU workload needs a GPU node; a latency-sensitive service might want to stay in one availability zone. Node affinity lets a pod's spec require or merely prefer nodes matching certain labels.

affinity:
  nodeAffinity:
    requiredDuringSchedulingIgnoredDuringExecution:
      nodeSelectorTerms:
        - matchExpressions:
            - key: gpu
              operator: In
              values: ["true"]

Taints & tolerations — the opposite direction

Affinity is a pod saying "I want that node." A taint is a node saying "stay away unless you're specifically allowed" — control-plane nodes are taints by default, which is why your regular app pods never accidentally land on them. A pod only schedules onto a tainted node if it carries a matching toleration.

# taint a node — nothing schedules here unless it tolerates this
kubectl taint nodes gpu-node-1 gpu=true:NoSchedule

# in the pod spec — the "permission slip" that lets it land there
tolerations:
  - key: gpu
    operator: Equal
    value: "true"
    effect: NoSchedule