Skip to main content

HPA Auto Scaling

Kubernetes-only feature

HPA (Horizontal Pod Autoscaler) is available only on the Kubernetes runtime. Docker · Podman runtimes have no auto-scaling or manual scaling features.

Pod List Tab (Kubernetes)

Manually adjusting scale every time traffic surges is tedious and risky. With HPA (Horizontal Pod Autoscaler), Pods automatically scale up or down based on service load.

Why Use HPA?
  • Automatic response: Pods are added automatically when traffic increases.
  • Cost savings: Reduces Pods when traffic is low to save resources.
  • Improved stability: Prevents manual operation mistakes.

HPA Capabilities in KIOPS​

KIOPS's HPA supports CPU-based auto-scaling only, created internally via kubectl autoscale --cpu-percent. Memory-based scaling, multiple metrics, custom metrics (such as Prometheus), scale stabilization windows (stabilizationWindowSeconds), and scaling policies (Pods/Percent type, value, period) are not supported.

The configurable fields are:

  • minReplicas: Minimum Pod count (default 1, recommended 2 or more)
  • maxReplicas: Maximum Pod count (default 10)
  • targetCPU(%): Target CPU utilization (default 80)

How It Works​

[Metrics Server] → [HPA Controller] → [Deployment Scale]
        ↓                   ↓                    ↓
   Collect CPU       Compare to target     Adjust Pod count

Prerequisites​

  • Metrics Server must be installed on the Kubernetes cluster.
  • The service must be deployed on K8s.

If Metrics Server is not installed, refer to the Metrics Server Installation guide.

Permission Notice: If you cannot access this feature, please request permission from your organization manager.

Creating an HPA​

Step 1: Open the Operations Modal​

  1. From the [Service Management] page, click the Operate stage for the target service.
  2. In the operations modal, go to the Pod List tab.

Step 2: Enter the HPA Section​

  1. Find the HPA section in the Pod List tab.
  2. Click the Create HPA button.

Step 3: Enter Settings​

Only the following three fields are required.

  • minReplicas (default 1, recommended 2 or more): Minimum Pods to maintain even under low load. Recommended at least 2 for high availability.
  • maxReplicas (default 10): Maximum Pods to scale up to. Set considering cluster resources and costs.
  • targetCPU(%) (default 80): Pods are added when CPU utilization exceeds this value. 70–80% is typically recommended.

Step 4: Create​

  1. Click the Create button.
  2. KIOPS creates the HPA via kubectl autoscale --cpu-percent.
  3. If an HPA already exists on the same Deployment, the existing HPA is deleted and recreated.

HPA Card Information​

A created HPA is displayed as a card in the Pod List tab.

  • HPA name
  • Target Deployment
  • Min / Max Pods
  • Target CPU(%)
  • Currently running Pod count
  • Delete button

Clicking the Delete button removes the HPA and reverts to manual scaling.

HPA vs. Manual Scaling​

While an HPA is active, manual scaling is locked. The backend also blocks kubectl scale in this state. To use manual scaling again, first remove the HPA via the Delete button on the HPA card.

Practical Use Scenarios​

Scenario 1: Web Application​

Requirements: Auto scale based on traffic.

Settings:

  • minReplicas: 2
  • maxReplicas: 20
  • targetCPU: 70%

Scenario 2: API Server​

Requirements: Fast scale up and stable operation.

Settings:

  • minReplicas: 3
  • maxReplicas: 50
  • targetCPU: 60%

Scenario 3: Batch Workload​

Requirements: Gradual scaling based on CPU load.

Settings:

  • minReplicas: 1
  • maxReplicas: 10
  • targetCPU: 80%

Resource Request Settings​

For HPA to work accurately, the Pod's CPU request (requests.cpu) must be configured. HPA calculates utilization based on this value when deciding to scale.

Resource Request Example​

resources:
  requests:
    cpu: "200m"
    memory: "256Mi"
  limits:
    cpu: "500m"
    memory: "512Mi"

With this configuration, HPA targetCPU 70% means:

  • 200m × 70% = 140m average usage maintains the current scale.
  • Above that, scale up is triggered.

Memory limit is used only to prevent OOM and is not used as a scaling criterion by HPA.

Troubleshooting​

HPA Not Scaling​

  • Metrics not displayed: Metrics Server is not installed. Install it from the Metrics Server Installation guide.
  • Status shown as Unknown: Check whether requests.cpu is configured on the Pod. HPA cannot calculate utilization without it.
  • Target not reached: Current CPU load is low and scaling is not needed. This is normal.

Excessive Scaling​

  • Too many Pods created: targetCPU is set too low. Increase to 70–80%.
  • Pods in Pending due to resource shortage: maxReplicas reached or cluster resources are insufficient. Adjust maxReplicas or add nodes.

When You Need to Delete HPA​

To use manual scaling or change settings, use the Delete button on the HPA card. Creating a new HPA on the same Deployment automatically deletes and recreates the existing HPA.

Best Practices​

Here are recommendations for using HPA effectively.

Configuration Recommendations​

Recommended Values
  1. minReplicas: 2 or more - Essential for high availability.
  2. targetCPU: 70–80% - Too low causes unnecessary costs.
  3. CPU request (requests.cpu): Based on actual usage - Required for HPA to work accurately.

Monitoring​

Regularly verify that HPA is working as intended.

  1. Check current Pod count: View the HPA card in the Pod List tab.
  2. Alert when max is reached: If maxReplicas is hit frequently, raise the limit or review resource requests.
  3. Cost monitoring: Track cloud costs associated with scaling.

Testing​

Validate HPA behavior before production deployment.

  1. Load testing: Increase traffic to verify scale up works correctly.
  2. Recovery testing: Reduce traffic to verify scale down works appropriately.
  3. Limit testing: Check service behavior when maxReplicas is reached.
Test Before Production

Don't apply HPA settings directly to production. Test thoroughly in a staging environment first.

Disabling HPA​

To remove an HPA, click the Delete button on its card in the Pod List tab. After deletion, manual scaling (Apply Pod Count / Remove All Pods) is available again.