HPA.md

Horizontal pod auto-scaling (HPA)

If Pods starts to use more Memory or CPU than defined in the HPA, Auto-scaling action will trigger and it will add another replica of the same deployment to balance the load.

Prerequisites

To make HPA work we need CPU and memory utilised by the pods. For this metrics server must be deployed.

Configuration

File: clutser.yaml

Add the following in cluster spec.

These lines will inform the kube controller manager not to use Rest based endpoints to gather metrics.

  kubeControllerManager:
    horizontalPodAutoscalerUseRestClients: false

File: site/dep.yaml

1. Add container usage limit. Metrics server will be tracking this usage limits. Eg. 50% of the requested resources.

         containers:
         - name: php-fpm
           image: superuser/php:v1.0.8
           resources:
             limits:
               cpu: "0.2"
               memory: 100Mi
             requests:
               cpu: 10m
               memory: 64Mi

2. Add an HPA block same as example :

   apiVersion: autoscaling/v2beta1
   kind: HorizontalPodAutoscaler
   metadata:
     name: test-aws-k8s-master
   spec:
     scaleTargetRef:
       apiVersion: apps/v1
       kind: Deployment
       name: test-aws-k8s-master #name of deployment
     minReplicas: 3 #maximum limit pods can replicate to
     maxReplicas: 1 #minimum number of replicas: >=1
     metrics:
     - type: Resource
       resource:
         name: memory #target resource
         targetAverageUtilization: 50 #trigger if utilization is >=50%
     - type: Resource
       resource:
         name: cpu
         targetAverageUtilization: 50

Limits

If Pod/Node affinities are there then HPA behaviour will depend on it.

1. If pod afinity is there and 2 nodes are available. If auto scaling triggers then maximum 2 numbers of replicas can be created.

2. If node afinity is there and only 2 nodes are available in the node pool then numer of replicas are depended on the resources are available in the nodepool.

3. If Pod and Node afinities are configured at a same time then both limits 1 and 2 will be applied.