HorizontalPodAutoscaler

Workloads / Guide 33 of 103 / Intermediate

• Kubernetes1.36.4
• Cluster4 nodes
• Runtimecontainerd 2.2.6
• CNICalico v3.32.1
• TimeAbout 35 min

Needs metrics-server. Without it the HPA reports FailedGetResourceMetric and never scales.

• APIautoscaling/v2

This guide includes

Use this when load varies and the replica count should follow it. This matters because an HPA computes a percentage of the CPU request rather than of the node, and its scale-down is deliberately slow - both of which surprise people watching one for the first time.

• creating an HPA with kubectl autoscale, and noting the flag it warns is on its way out
• reading back the API version the object was actually stored as
• seeing the error it reports for the first minute, before any metrics have arrived
• putting real load on with a tight request loop and watching replicas climb
• removing the load and finding scale-down deliberately slow, governed by a stabilisation window

Before you start

• The metrics-server guide. The HPA reads the same Metrics API kubectl top does.
• The requests, limits and QoS guide, because the target is a percentage of the request.
• The Deployments guide.

1. 1

   Create it, and note the flag that is on its way out

2. 2

   Idle, and the error you see for the first minute

3. 3

   Real load, and the scale-up

4. 4

   Remove the load: scale-down is deliberately slow

Official sources

• Horizontal Pod Autoscaling
• HorizontalPodAutoscaler walkthrough
• Configurable scaling behavior
• Resource metrics pipeline
• HorizontalPodAutoscaler API