Skip to content
Branch: master
Find file History
Latest commit 0f34ed7 Jul 3, 2019

Custom Metrics - Stackdriver Adapter

Custom Metrics - Stackdriver Adapter is an implementation of Custom Metrics API and External Metrics API using Stackdriver as a backend. Its purpose is to enable pod autoscaling based on Stackdriver custom metrics.

Usage guide

This guide shows how to set up Custom Metrics - Stackdriver Adapter and export metrics to Stackdriver in a compatible way. Once this is done, you can use them to scale your application, following HPA walkthrough.

Configure cluster

  1. Create Kubernetes cluster or use existing one, see cluster setup. Requirements:

    • Kubernetes version 1.8.1 or newer running on GKE or GCE

    • Monitoring scope monitoring set up on cluster nodes. It is enabled by default, so you should not have to do anything. See also OAuth 2.0 API Scopes to learn more about authentication scopes.

      You can use following commands to verify that the scopes are set correctly:

      • For GKE cluster <my_cluster>, use following command:
        gcloud container clusters describe <my_cluster>
        For each node pool check the section oauthScopes - there should be scope listed there.
      • For a GCE instance <my_instance> use following command:
        gcloud compute instances describe <my_instance>
        should be listed in the scopes section.

      To configure set scopes manually, you can use:

      • --scopes flag if you are using gcloud container clusters create command, see gcloud documentation.
      • Environment variable NODE_SCOPES if you are using script. It is enabled by default.
      • To set scopes in existing clusters you can use gcloud beta compute instances set-scopes command, see gcloud documentation.
    • On GKE, you need cluster-admin permissions on your cluster. You can grant your user account these permissions with following command:

      kubectl create clusterrolebinding cluster-admin-binding --clusterrole cluster-admin --user $(gcloud config get-value account)
  2. Start Custom Metrics - Stackdriver Adapter.

Stackdriver supports two models of Kubernetes resources: the legacy model using monitored resource gke_container and the new model using different Kubernetes monitored resources, including for example k8s_pod, k8s_node. See monitored resources documentation for more details.

  • If you use legacy resource model:
    kubectl apply -f
  • If you use new resource model:
    kubectl apply -f

Metrics available from Stackdriver

Custom Metrics - Stackdriver Adapter exposes Stackdriver metrics to Kubernetes components via two APIs.

  1. Any Stackdriver metric can be retrieved via External Metrics API with one assumption: metricType = DOUBLE or INT64. For example, this API can be used to configure Horizontal Pod Autoscaler to scale deployment based on any of existing metrics from other GCP services.

  2. Metrics attached to Kubernetes objects, such as Pod or Node, can be retrieved via Custom Metrics API. The following section provides more details about exporting such metrics.

Metric kinds

Stackdriver specifies three metric kinds, all of which are supported by Custom Metrics - Stackdriver Adapter:

  1. GAUGE - Each data point represents an instantaneous measurement, for example the temperature. The adapter exposes the latest value.
  2. DELTA - Each data point represents the change in a value over the time interval. The adapter exposes rate of the metric - the metric change per second computed over last 5 minutes.
  3. CUMULATIVE - Each data point is a value being accumulated over time. The adapter exposes rate of the metric - the metric change per second computed over last 5 minutes.

Export custom metrics to Stackdriver

To learn how to create your custom metric and write your data to Stackdriver, follow Stackdriver custom metrics documentation. You can also follow Prometheus to Stackdriver documentation to export metrics exposed by your pods in Prometheus format.

The name of your metric must start with prefix followed by a simple name, as defined in custom metric naming rules.

You will report your metric against a appropriate monitored resource for Kubernetes objects. To use legacy resource model, use monitored resource gke_container. To use new resource model, use one of monitored resources: k8s_pod or k8s_node - corresponding to Kubernetes objects Pod and Node.

  1. Define your custom metric by following Stackdriver custom metrics documentation. Your metric descriptor needs to meet following requirements:
    • metricType = DOUBLE or INT64
  2. Export metric from your application. The metric has to be associated with a specific pod or node and meet folowing requirements:
    • resource_type set accordingly: gke_container for legacy resource model or one of k8s_pod, k8s_node for new resource model. (See monitored resources documentation)

    • All resource labels for your monitored resource set to correct values. In particular:

      • pod_id, pod_name, namespace_name can be obtained via downward API. Example configuration that passes these values to your application as flags:

        apiVersion: v1
        kind: Pod
          name: my-pod
          - image: <my-image>
            - my-app --pod_id=$(POD_ID) --pod_name=$(POD_NAME) --namespace_name=$(NAMESPACE_NAME)
            - name: POD_ID
                  fieldPath: metadata.uid
            - name: POD_NAME
            - name: NAMESPACE_NAME
                  fieldPath: metadata.namespace

        Example flag definition in Go:

        import "flag"
        podIdFlag := flag.String("pod_id", "", "a string")
        podID := *podIdFlag
      • For monitored resource gke_container, container_name should be set to "" to indicate that the metric is associated with a pod, not a particular container.

      • project_id, zone, location, cluster_name - can be obtained by your application from [metadata server]. You can use Google Cloud compute metadata client to get these values, example in Go:

        import gce ""
        project_id, err := gce.ProjectID()
        // the zone where your application runs
        // used in legacy resource model
        zone, err := gce.Zone()
        // cluster location can be different than your application zone in
        // clusters spanning across multiple zones
        // used in new resource model
        location, err := gce.InstanceAttributeValue("cluster-location")
        cluster_name, err := gce.InstanceAttributeValue("cluster-name")
      • namespace_id and instance_id (for legacy resource model) are not used by Custom Metrics - Stackdriver Adapter, but still it's recommended to set those to the correct values to make them more useful for other use cases.

      Example code exporting a metric to Stackdriver, written in Go:

      import (
      // Create stackdriver client
      authClient := oauth2.NewClient(context.Background(), google.ComputeTokenSource(""))
      stackdriverService, err := v3.New(oauthClient)
      if err != nil {
      // Define metric time series filling in all required fields
      request := &v3.CreateTimeSeriesRequest{
        TimeSeries: []*v3.TimeSeries{
            Metric: &v3.Metric{
              Type: "" + <your metric name>,
            Resource: &v3.MonitoredResource{
              Type: "k8s_pod",
              Labels: map[string]string{
                "project_id":     <your project ID>,
                "location":       <your cluster location>,
                "cluster_name":   <your cluster name>,
                "namespace_name": <your namespace>,
                "pod_name":       <your pod name>,
            Points: []*v3.Point{
                Interval: &v3.TimeInterval{
                  EndTime: time.Now().Format(time.RFC3339),
                Value: &v3.TypedValue{
                  Int64Value: <your metric value>,
      stackdriverService.Projects.TimeSeries.Create("projects/<your project ID>", request).Do()


To test your custom metrics setup or see a reference on how to push your metrics to Stackdriver, check out our examples:

You can’t perform that action at this time.