Kubernetes Horizontal Pod Autoscaler for Hazelcast

Deploy Hazelcast on a Kubernetes cluster and set up Kubernetes Horizontal Pod Autoscaler (HPA), using Resource Metrics and Custom Metrics.

Before you Begin

  • Kubernetes cluster with kubectl configured

  • Helm 3 tool installed

Install Hazelcast Helm Chart

Helm is the package manager for Kubernetes, and we will use it throughout the document to install various software. This is the link to install helm into your computer. Once you have a working Helm 3 we can install Hazelcast Helm Chart by running the following commands:

$ helm repo add hazelcast https://hazelcast-charts.s3.amazonaws.com/
$ helm repo update
$ helm install hazelcast hazelcast/hazelcast

Now you can verify that Hazelcast cluster is successfully deployed:

$ kubectl get all
NAME                        READY   STATUS    RESTARTS   AGE
pod/hazelcast-0             1/1     Running   0          7m35s
pod/hazelcast-1             1/1     Running   0          7m3s
pod/hazelcast-2             1/1     Running   0          6m23s
pod/hazelcast-mancenter-0   1/1     Running   0          7m35s

NAME                          TYPE           CLUSTER-IP      EXTERNAL-IP    PORT(S)                        AGE
service/hazelcast             ClusterIP      None            <none>         5701/TCP                       7m35s
service/hazelcast-mancenter   LoadBalancer   8080:30957/TCP,443:30042/TCP   7m35s
service/kubernetes            ClusterIP    <none>         443/TCP                        25h

NAME                                   READY   AGE
statefulset.apps/hazelcast             3/3     7m35s
statefulset.apps/hazelcast-mancenter   1/1     7m35s

For more details on installing and using Hazelcast Helm Chart you can see this blogpost.

Horizontal Pod AutoScaler (Resource Metrics)

As you can read the details from official Kubernetes documentation, Resource Metrics provide CPU and Memory based metrics for pods and nodes in your Kubernetes Cluster. Those metrics are exposed via the metrics.k8s.io API and one implementation of that API is Metrics Server.

Before moving forward, verify that Metrics Server is properly installed and visible in the list of API Registration.

$ kubectl get apiservices.apiregistration.k8s.io | grep metrics-server
v1beta1.metrics.k8s.io                 kube-system/metrics-server   True        26h

Let’s create an HPA based on CPU usage.

$ kubectl autoscale statefulset hazelcast --cpu-percent=50 --min=3 --max=10

This HPA will periodically check Hazelcast StatefulSet CPU usage and will decide the number of running pods between 3 and 10, based on some calculations. The simplest way to put some cpu load on Hazelcast Pod is to execute yes tool. This is just to show how HPA is triggered to scale up Hazelcast Cluster by printing yes in one of Hazelcast pods. You should use a proper load testing tool to test HPA in your Hazelcast Cluster. Before generating CPU load, you can open 2 new terminals to watch HPA target values and number of Hazelcast pods via watch kubectl get pods and watch kubectl get hpa commands. Let’s move on and execute the following command for 5-10 seconds and terminate via Ctrl + C.

Now that your HPA target is above 50%, you should see some new pods have been created. As the initial Hazelcast cluster was 3 members cluster, hazelcast-3 and above are the new pods created by HPA.

$ kubectl get hpa
NAME             REFERENCE                  TARGETS     MINPODS   MAXPODS   REPLICAS   AGE
hazelcast        StatefulSet/hazelcast      282m/200m   3         5         5          18m

$ kubectl get pods
NAME                                    READY   STATUS        RESTARTS   AGE
hazelcast-0                             1/1     Running       0          30m
hazelcast-1                             1/1     Running       0          30m
hazelcast-2                             1/1     Running       0          29m
hazelcast-3                             1/1     Running       0          14m
hazelcast-4                             1/1     Running       0          13m
hazelcast-mancenter-0                   1/1     Running       1          30m

Clean Up

As you successfully managed to use Resource Metrics with Hazelcast, we can clean up resources used up to that point.

$ helm uninstall hazelcast
release "hazelcast" uninstalled

$ kubectl delete hpa hazelcast
horizontalpodautoscaler.autoscaling "hazelcast" deleted

Custom Metrics

In the previous section, we have explained how to use Resource Metrics to autoscale your deployments based on CPU or Memory Metrics. Although that is fine for some architectures, those metrics are Kubernetes Pod or Node level so application-level autoscaling is not possible with Resource Metrics. Kubernetes introduced Custom Metrics API in order to fill in this gap. When using Custom Metrics API, each container exposes its own metrics and HPA uses those metrics to make autoscaling decisions. In this example, we will use Prometheus as Metrics Storage and Prometheus Adapter as the Custom Metrics API Provider.

Install Prometheus and Prometheus Adapter

To install Prometheus, execute the following commands:

$ helm repo add prometheus-community https://prometheus-community.github.io/helm-charts
"prometheus-community" has been added to your repositories

$ kubectl create ns monitoring
namespace/monitoring created

$ helm install prometheus prometheus-community/prometheus --namespace=monitoring

Before installing Prometheus Adapter let’s take a look at the content of prometheus-adapter-values.yaml

  default: true
  - seriesQuery: '{__name__=~"jvm_memory_bytes_(used|max)",area="heap",kubernetes_name=~"hazelcast.*"}'
    - is: ^jvm_memory_bytes_(used|max)$
        kubernetes_pod_name: {resource: "pod"}
        kubernetes_namespace: {resource: "namespace"}
        kubernetes_name: {resource: "service"}
      matches: ^jvm_memory_bytes_(used|max)$
      as: "on_heap_ratio"
    metricsQuery: max(jvm_memory_bytes_used{<<.LabelMatchers>>}/jvm_memory_bytes_max{<<.LabelMatchers>>}) by (<<.GroupBy>>)
  url: http://prometheus-server # make sure the url is correct
  port: 80

This configuration will be passed to the helm chart while deploying Prometheus Adapter but let’s just go through each part before doing that. The config basically tells Prometheus Adapter:

  • query only jvm_memory_bytes_used and jvm_memory_bytes_max

  • assign kubernetes_* based labels to resources to be able to query via REST URLs like /apis/custom.metrics.k8s.io/v1beta1/namespaces/default/services/*/on_heap_ratio

  • give a new, easier name (on_heap_ratio) to the metric that we expose via custom metrics adapter

  • select max value out of all series provided by all PODs

This example uses max function while creating metricsQuery, but you can basically use some other aggregation operator like avg in your own configuration. If you saved the file above, you can create a prometheus adapter based on that configuration.

To install Prometheus Adapter run the following command:

$ helm install prometheus-adpter prometheus-community/prometheus-adapter --namespace=monitoring -f prometheus-adapter-values.yaml

Install Metrics Enabled Hazelcast Cluster

Let’s install a new 3 members Hazelcast cluster with metrics enabled. Each Hazelcast member container in this new deployment will expose their own metrics data under /metrics endpoint. This endpoint exposes metrics in Prometheus format because each Hazelcast container is started with Prometheus JMX Exporter. This is a feature provided by Hazelcast Docker Image. We also set resources.limits.memory=512Mi which sets each Hazelcast member JVM max heap size to 128Mi. JVM by default grabs 25% of available memory as max heap size.

$ helm install hazelcast hazelcast/hazelcast --set metrics.enabled=true,resources.limits.memory=512Mi

Verify that the custom rule we provided to Prometheus Adapter is functioning properly. If you see "Error from server (NotFound): the server could not find the metric on_heap_ratio for services", you might need to wait some time because Prometheus might not have started scraping Hazelcast specific metrics.

$ kubectl get --raw "/apis/custom.metrics.k8s.io/v1beta1/namespaces/default/services/*/on_heap_ratio" | jq .
  "kind": "MetricValueList",
  "apiVersion": "custom.metrics.k8s.io/v1beta1",
  "metadata": {
    "selfLink": "/apis/custom.metrics.k8s.io/v1beta1/namespaces/default/services/%2A/on_heap_ratio"
  "items": [
      "describedObject": {
        "kind": "Service",
        "namespace": "default",
        "name": "hazelcast-metrics",
        "apiVersion": "/v1"
      "metricName": "on_heap_ratio",
      "timestamp": "2020-11-13T12:57:23Z",
      "value": "108m",
      "selector": null

The most important part in this output is "value": 108m. The suffix m means milli-unit, and it is kubernetes-style quantities to define metric values. Milli-unit is equivalent to 1000ths of a unit so 108m is actually referring to 3.3% which means max value of on_heap_ratio seen so far.

Install Horizontal Pod AutoScaler (Custom Metrics)

As we have configured Hazelcast, Prometheus and Prometheus Adapter, let’s create a Horizontal Pod AutoScaler based on on_heap_ratio metric. hazelcast-custom-metrics-hpa.yaml tells HPA if targetValue > 200m then scale up my cluster. 200m as we explained above means actually 20%. You can change that number based on your own use case.

See the content of hazelcast-custom-metrics-hpa.yaml file.

apiVersion: autoscaling/v2beta2
kind: HorizontalPodAutoscaler
  name: heap-based-hpa
    apiVersion: apps/v1
    kind: StatefulSet
    name: hazelcast
  minReplicas: 3
  maxReplicas: 10
    - type: Object
          kind: Service
          name: hazelcast-metrics
          name: on_heap_ratio
          type: Value
          value: 200m

Apply HPA to your cluster with kubectl:

$ kubectl apply -f hazelcast-custom-metrics-hpa.yaml
horizontalpodautoscaler.autoscaling/heap-based-hpa created

Generate some Memory Load for HPA

Let’s have a look TARGETS part of HPA output

$ kubectl get hpa heap-based-hpa
heap-based-hpa   StatefulSet/hazelcast   180m/200m   3         10        10         11m

As we see, current HPA Target is 180m/200m so if we increase memory usage just 10% by adding 10MB into the cluster, HPA should trigger a scale up event. I will use Hazelcast Java Client to put some data into cluster, but you can use your own language to implement the same functionality. You can see all Hazelcast supported programming languages here.

Let’s first port forward from our local machine to be able to connect remote k8s Hazelcast member pod.

$ kubectl port-forward hazelcast-0 5701

Execute following code snippet to put data into Hazelcast Cluster.

 // start Hazelcast Client with smartRouting disabled
ClientConfig cfg = new ClientConfig();
HazelcastInstance client = HazelcastClient.newHazelcastClient(cfg);
// create Hazelcast Distributed Map "numbers"
IMap<Object, Object> numbers = client.getMap("numbers");

// put 10000*1K = 10M to "numbers"
int i=0;
while (i++ < 10000)
    numbers.put(i,new byte[1024]);

// check the size of "numbers"

//clean up

You can use the client application from the External Smart Client tutorial to connect to the Hazelcast cluster.

When you start putting data into your Hazelcast cluster, you will see that new pods will be created and added to Hazelcast Cluster.

$ kubectl get po
NAME                    READY   STATUS    RESTARTS   AGE
hazelcast-0             1/1     Running   0          73m
hazelcast-1             1/1     Running   0          72m
hazelcast-2             1/1     Running   0          72m
hazelcast-3             1/1     Running   0          16m
hazelcast-4             1/1     Running   0          3m
hazelcast-5             1/1     Running   0          3m
hazelcast-6             1/1     Running   0          47s
hazelcast-7             1/1     Running   0          41s
hazelcast-8             1/1     Running   0          14s
hazelcast-9             1/1     Running   0          13s
hazelcast-mancenter-0   1/1     Running   0          73m


Autoscaling is an important feature for enterprises to save money and to cope with unexpected traffic coming to your deployments. However, configuring autoscaling needs to be done carefully because you might end up unnecessary scale up/down operations which might cost some instability in your system. In this guide, we explained how you can use HPA with your Hazelcast Cluster based on Resource Metrics and Custom Metrics. If Kubernetes Pod/Node Level Cpu/Memory usage is fine for you then use Resource Metrics. If you have more specific requirements, and you need to have Hazelcast specific autoscaling capabilities, Custom Metrics is the answer.

Software Versions

This is the list of software versions used in this guide.

$ helm ls --all-namespaces
NAME             	NAMESPACE 	REVISION	UPDATED                             	STATUS  	CHART                   	APP VERSION
hazelcast        	default   	1       	2020-11-13 14:53:37.702621 +0200 EET	deployed	hazelcast-3.5.0         	4.1
prometheus       	monitoring	1       	2020-11-13 14:35:50.630099 +0200 EET	deployed	prometheus-11.16.8      	2.21.0
prometheus-adpter	monitoring	1       	2020-11-13 14:49:11.954809 +0200 EET	deployed	prometheus-adapter-2.7.0	v0.7.0