Horizontal Pod Autoscaler
The HPA automatically adds or removes Pod replicas based on observed CPU, memory, or custom metrics.
Traffic rises and falls. Manually editing replicas is slow and wasteful. The Horizontal Pod Autoscaler (HPA) scales a Deployment automatically to match load.
How it decides
You give the HPA a target, such as "keep average CPU at 50%". It watches the metric and:
- Adds Pods when the metric is above target.
- Removes Pods when it is comfortably below.
You also set a minReplicas and maxReplicas so it never scales to zero or to infinity.
Prerequisites
The HPA needs a metrics source (the Metrics Server for CPU/memory), and your Pods must declare resource requests — the percentage is measured against them.
Example
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
name: web-hpa
spec:
scaleTargetRef:
apiVersion: apps/v1
kind: Deployment
name: web
minReplicas: 2
maxReplicas: 10
metrics:
- type: Resource
resource:
name: cpu
target:
type: Utilization
averageUtilization: 50When to use it
- An e-commerce site uses HPA so its product-search API automatically scales from 3 to 20 replicas during flash sale traffic spikes.
- A data processing service uses HPA with custom metrics (queue depth) to add workers when the message queue grows beyond 100 items.
- A team reduces cloud costs by setting HPA minReplicas=2 so the cluster scales down idle deployments overnight without manual intervention.
More examples
HPA based on CPU utilization
Creates an HPA that keeps average CPU utilization at 60%, scaling the web deployment between 2 and 20 replicas.
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
metadata:
name: web-hpa
spec:
scaleTargetRef:
apiVersion: apps/v1
kind: Deployment
name: web
minReplicas: 2
maxReplicas: 20
metrics:
- type: Resource
resource:
name: cpu
target:
type: Utilization
averageUtilization: 60Inspect HPA scaling events
Shows current CPU utilization vs. target and the scaling events that Kubernetes has taken to maintain the target.
kubectl get hpa web-hpa
# NAME REFERENCE TARGETS MINPODS MAXPODS REPLICAS
# web-hpa Deployment/web 85%/60% 2 20 5
kubectl describe hpa web-hpa | grep -A5 'Events:'Scale down behaviour tuning
Configures a 5-minute stabilization window and limits scale-down to 25% of pods per minute to prevent thrashing.
spec:
behavior:
scaleDown:
stabilizationWindowSeconds: 300
policies:
- type: Percent
value: 25
periodSeconds: 60
Discussion