custom-autoscaler

OpenShift Custom Autoscaler Quickstart

The following illustrates how one might use OpenShift's metrics stack (Hawkular, Cassandra, Heapster) to develop customized application autoscaling.

Running & Testing

This script makes a few assumptions about your application

• The pods are managed by a deploymentConfig
• The deploymentConfig is named the same as the app. (in this case, myapp)
• The deploymentConfig has memory limits set on the containers. i.e.

  apiVersion: v1
  kind: DeploymentConfig
  metadata:
    name: myapp
  spec:
    template:
      spec:
        container:
        - resources:
            limits:
              memory: 140Mi
  
  

Before running the script, you need to export a few variables into your environment. The following represents a minimal setup. A full description of all variables is below.

export NAMESPACE=myproject
export TOKEN=`oc whoami -t`
export APP_NAME=myapp
export HAWKULAR_HOSTNAME=hawkular-metrics.apps.ocp-c1.myorg.com

From there we can simply run our script.

./scale.py

The script will continue to run until exited, either through Ctrl+C or killing the pid.

Environment Variables

Variable Name	Default Value	Description
NAMESPACE	n/a; Required	Namespace the app is running in.
TOKEN	n/a; Required	OpenShift User or ServiceAccount token. Can be retrieved using oc whoami -t for users or oc serviceaccount get-token <sa name> for serviceaccounts.
APP_NAME	n/a; Required	Name of DeploymentConfig. Must match name shown in oc get dc
HAWKULAR_HOSTNAME	n/a; Required	Route hostname for hawkular service. i.e. oc get route hawkular-metrics -n openshift-infra
METRIC_PULL_INTERVAL_SECONDS	60	Number of seconds between each scale up check
SCALE_UP_THRESHOLD	.7	Floating numeric value 0-1; Percentage of total memory used over which a scale-up will happen
CA_CERT	False	Path to a CA Certificate file to validate Hawkular certificate. Default value of False will turn off certificate checks and flash a warning message.

Issues, Limitations & Future Enhancements

• Solution should be deployed as a pod in openshift.
• Currently only scales up, not down
• Implementation very specific to memory, but could probably be genericized to pull some other metric.
• Needs better error handling