Skip to main content
Version: Next

FederatedHPA scales with custom metrics

In Karmada, a FederatedHPA scales up/down the workload's replicas across multiple clusters, with the aim of automatically scaling the workload to match the demand.

FederatedHPA not only supports resource metrics such as CPU and memory, but also supports custom metrics which may expand the use cases of FederatedHPA.

This document walks you through an example of enabling FederatedHPA to automatically manage scale for a cross-cluster app with custom metrics.

The walkthrough example will do as follows:

  • One sample-deployment's pod exists in member1 cluster.
  • The service is deployed in member1 and member2 cluster.
  • A MultiClusterService enables cross-cluster access between the sample-app services in member1 and member2.
  • Request the service from member1, route traffic to backend pods in both clusters, and trigger an increase in the pods' custom metrics(http_requests_total).
  • The replicas will be scaled up in member1 and member2 cluster.

In this example, the role of MultiClusterService is to load balance user requests to different member clusters, simulating a real-world scenario where applications in multiple clusters share the load. In practice, MultiClusterService is not required.

Prerequisites​

Karmada has been installed​

You can install Karmada by referring to Quick Start, or directly run hack/local-up-karmada.sh script which is also used to run our E2E cases.

Member Cluster Network​

Ensure that at least two clusters have been added to Karmada, and the container networks between member clusters are connected.

  • If you use the hack/local-up-karmada.sh script to deploy Karmada, Karmada will have three member clusters, and the container networks of the member1 and member2 will be connected.
  • You can use Submariner or other related open source projects to connect networks between member clusters.

Note: In order to prevent routing conflicts, Pod and Service CIDRs of clusters need non-overlapping.

prometheus and prometheus-adapter have been installed in member clusters​

You need to install prometheus and prometheus-adapter for member clusters to provide the custom metrics. You can install it by running the following in member clusters:

git clone https://github.com/prometheus-operator/kube-prometheus.git
cd kube-prometheus
kubectl apply --server-side -f manifests/setup
kubectl wait \
	--for condition=Established \
	--all CustomResourceDefinition \
	--namespace=monitoring
kubectl apply -f manifests/

You can verify the installation by the following command:

$ kubectl --kubeconfig=/root/.kube/members.config --context=member1 get po -nmonitoring
NAME                                   READY   STATUS    RESTARTS   AGE
alertmanager-main-0                    2/2     Running   0          30h
alertmanager-main-1                    2/2     Running   0          30h
alertmanager-main-2                    2/2     Running   0          30h
blackbox-exporter-6bc47b9578-zcbb7     3/3     Running   0          30h
grafana-6b68cd6b-vmw74                 1/1     Running   0          30h
kube-state-metrics-597db7f85d-2hpfs    3/3     Running   0          30h
node-exporter-q8hdx                    2/2     Running   0          30h
prometheus-adapter-57d9587488-86ckj    1/1     Running   0          29h
prometheus-adapter-57d9587488-zrt29    1/1     Running   0          29h
prometheus-k8s-0                       2/2     Running   0          30h
prometheus-k8s-1                       2/2     Running   0          30h
prometheus-operator-7d4b94944f-kkwkk   2/2     Running   0          30h

karmada-metrics-adapter has been installed in Karmada control plane​

You need to install karmada-metrics-adapter in Karmada control plane to provide the metrics API, install it by running:

hack/deploy-metrics-adapter.sh ${host_cluster_kubeconfig} ${host_cluster_context} ${karmada_apiserver_kubeconfig} ${karmada_apiserver_context_name}

If you use the hack/local-up-karmada.sh script to deploy Karmada, karmada-metrics-adapter will be installed by default.

Deploy workload in member1 and member2 cluster​

You need to deploy a sample deployment (1 replica) and service in member1 and member2.

apiVersion: apps/v1
kind: Deployment
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  replicas: 1
  selector:
    matchLabels:
      app: sample-app
  template:
    metadata:
      labels:
        app: sample-app
    spec:
      containers:
      - image: luxas/autoscale-demo:v0.1.2
        name: metrics-provider
        ports:
        - name: http
          containerPort: 8080
---
apiVersion: v1
kind: Service
metadata:
  labels:
    app: sample-app
  name: sample-app
spec:
  ports:
  - name: http
    port: 80
    protocol: TCP
    targetPort: 8080
  selector:
    app: sample-app
  type: ClusterIP
---
apiVersion: policy.karmada.io/v1alpha1
kind: PropagationPolicy
metadata:
  name: app-propagation
spec:
  resourceSelectors:
    - apiVersion: apps/v1
      kind: Deployment
      name: sample-app
    - apiVersion: v1
      kind: Service
      name: sample-app
  placement:
    clusterAffinity:
      clusterNames:
        - member1
        - member2
    replicaScheduling:
      replicaDivisionPreference: Weighted
      replicaSchedulingType: Divided
      weightPreference:
        staticWeightList:
          - targetCluster:
              clusterNames:
                - member1
            weight: 1
          - targetCluster:
              clusterNames:
                - member2
            weight: 1

After deploying, you can check the distribution of the Pods and Service:

$ karmadactl get pods --operation-scope members
NAME                                  CLUSTER   READY   STATUS    RESTARTS   AGE
sample-app-9b7d8c9f5-xrnfx            member1   1/1     Running   0          111s
$ karmadactl get svc --operation-scope members
NAME                 CLUSTER   TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)   AGE     ADOPTION
sample-app           member1   ClusterIP   10.11.29.250    <none>        80/TCP    3m53s   Y

Monitor your application in member1 and member2 cluster​

In order to monitor your application, you'll need to set up a ServiceMonitor pointing at the application. Assuming you've set up your Prometheus instance to use ServiceMonitors with the app: sample-app label, create a ServiceMonitor to monitor the app's metrics via the service:

apiVersion: monitoring.coreos.com/v1
kind: ServiceMonitor
metadata:
  name: sample-app
  labels:
    app: sample-app
spec:
  selector:
    matchLabels:
      app: sample-app
  endpoints:
  - port: http
kubectl create -f sample-app.monitor.yaml

Now, you should see your metrics (http_requests_total) appear in your Prometheus instance. Look them up via the dashboard, and make sure they have the namespace and pod labels. If not, check the labels on the service monitor match the ones on the Prometheus CRD.

Launch you adapter in member1 and member2 cluster​

After you deploy prometheus-adapter, you need to update to the adapter config which is necessary in order to expose custom metrics.

apiVersion: v1
kind: ConfigMap
metadata:
  name: adapter-config
  namespace: monitoring
data:
  config.yaml: |-
    "rules":
    - "seriesQuery": |
         {namespace!="",__name__!~"^container_.*"}
      "resources":
        "template": "<<.Resource>>"
      "name":
        "matches": "^(.*)_total"
        "as": ""
      "metricsQuery": |
        sum by (<<.GroupBy>>) (
          irate (
            <<.Series>>{<<.LabelMatchers>>}[1m]
          )
        )
kubectl apply -f prom-adapter.config.yaml
# Restart prom-adapter pods
kubectl rollout restart deployment prometheus-adapter -n monitoring

Register metrics API in member1 and member2 cluster​

You also need to register the custom metrics API with the API aggregator (part of the main Kubernetes API server). For that you need to create an APIService resource.

apiVersion: apiregistration.k8s.io/v1
kind: APIService
metadata:
  name: v1beta2.custom.metrics.k8s.io
spec:
  group: custom.metrics.k8s.io
  groupPriorityMinimum: 100
  insecureSkipTLSVerify: true
  service:
    name: prometheus-adapter
    namespace: monitoring
  version: v1beta2
  versionPriority: 100
kubectl create -f api-service.yaml

The API is registered as custom.metrics.k8s.io/v1beta2, and you can use the following command to verify:

kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta2/namespaces/default/pods/*/http_requests?selector=app%3Dsample-app"

The output is similar to:

{
  "kind": "MetricValueList", 
  "apiVersion": "custom.metrics.k8s.io/v1beta2", 
  "metadata": {}, 
  "items": [
    {
      "describedObject": {
        "kind": "Pod", 
        "namespace": "default", 
        "name": "sample-app-9b7d8c9f5-9lw6b", 
        "apiVersion": "/v1"
      }, 
      "metric": {
        "name": "http_requests", 
        "selector": null
      }, 
      "timestamp": "2023-06-14T09:09:54Z", 
      "value": "66m"
    }
  ]
}

If karmada-metrics-adapter is installed successfully, you can also verify it with the above command in Karmada control plane.

Deploy FederatedHPA in Karmada control plane​

Then let's deploy FederatedHPA in Karmada control plane.

apiVersion: autoscaling.karmada.io/v1alpha1
kind: FederatedHPA
metadata:
  name: sample-app
spec:
  scaleTargetRef:
    apiVersion: apps/v1
    kind: Deployment
    name: sample-app
  minReplicas: 1
  maxReplicas: 10
  behavior:
    scaleDown:
      stabilizationWindowSeconds: 10
    scaleUp:
      stabilizationWindowSeconds: 10
  metrics:
    - type: Pods
      pods:
        metric:
          name: http_requests
        target:
          averageValue: 700m
          type: Value

After deploying, you can check the FederatedHPA:

NAME          REFERENCE-KIND   REFERENCE-NAME   MINPODS   MAXPODS   REPLICAS   AGE
sample-app    Deployment       sample-app       1         10        1          15d

Create a MultiClusterService for cross-cluster access​

The MultiClusterService is used to enable cross-cluster access between the sample-app services in member1 and member2. When a client in member1 accesses sample-app, the request is not limited to the backend pods local to member1; it can also be routed to backend pods behind the sample-app service in member2. More usage of MultiClusterService can be found in MultiClusterService user guide.

Create a MultiClusterService object on Karmada Control Plane.

apiVersion: networking.karmada.io/v1alpha1
kind: MultiClusterService
metadata:
  name: sample-app
spec:
  types:
    - CrossCluster
  consumerClusters:
    - name: member1
    - name: member2
  providerClusters:
    - name: member1
    - name: member2

After deploying, requests sent from member1 to sample-app can be routed to backend pods in both member1 and member2.

Install hey http load testing tool in member1 cluster​

In order to do http requests, here you can use hey.

  • Download hey and copy it to kind cluster container.
wget -O hey https://storage.googleapis.com/hey-releases/hey_linux_amd64
chmod +x hey
docker cp hey member1-control-plane:/usr/local/bin/hey

Test scaling up​

  • Check the pod distribution firstly.

    $ karmadactl get pods --operation-scope members
    NAME                                  CLUSTER   READY   STATUS    RESTARTS   AGE
    sample-app-9b7d8c9f5-xrnfx            member1   1/1     Running   0          111s

  • Check the service IP in member1.

    $ karmadactl get svc --operation-scope members
    NAME            CLUSTER   TYPE        CLUSTER-IP        EXTERNAL-IP   PORT(S)   AGE   ADOPTION
    sample-app      member1   ClusterIP   10.11.59.213      <none>        80/TCP    20m   Y

  • Request the service from member1 with hey to increase the sample-app pods' custom metric(http_requests_total) across both clusters.

    docker exec member1-control-plane hey -c 1000 -z 1m http://10.11.59.213/metrics

  • Wait 15s, the replicas will be scaled up, then you can check the pod distribution again.

    $ karmadactl get pods --operation-scope members -l app=sample-app
    NAME                         CLUSTER   READY   STATUS    RESTARTS   AGE
    sample-app-9b7d8c9f5-454vz   member2   1/1     Running   0          84s
    sample-app-9b7d8c9f5-7fjhn   member2   1/1     Running   0          69s
    sample-app-9b7d8c9f5-ddf4s   member2   1/1     Running   0          69s
    sample-app-9b7d8c9f5-mxqmh   member2   1/1     Running   0          84s
    sample-app-9b7d8c9f5-qbc2j   member2   1/1     Running   0          69s
    sample-app-9b7d8c9f5-2tgxt   member1   1/1     Running   0          69s
    sample-app-9b7d8c9f5-66n9s   member1   1/1     Running   0          69s
    sample-app-9b7d8c9f5-fbzps   member1   1/1     Running   0          84s
    sample-app-9b7d8c9f5-ldmhz   member1   1/1     Running   0          84s
    sample-app-9b7d8c9f5-xrnfx   member1   1/1     Running   0          87m

Test scaling down​

After 1 minute, the load testing tool will be stopped, then you can see the workload is scaled down across clusters.

$ karmadactl get pods --operation-scope members -l app=sample-app
NAME                         CLUSTER   READY   STATUS    RESTARTS   AGE
sample-app-9b7d8c9f5-xrnfx   member1   1/1     Running   0          91m