Deploying Pulsar on Kubernetes

Pulsar version 2.0

The documentation that you’re reading is for the 2.0 release of Apache Pulsar. For more information on Pulsar 2.0, see this guide.

Pulsar can be easily deployed in Kubernetes clusters, either in managed clusters on Google Kubernetes Engine or Amazon Web Services or in custom clusters.

The deployment method shown in this guide relies on YAML definitions for Kubernetes resources. The kubernetes subdirectory of the Pulsar package holds resource definitions for:


To get started, install a source package from the downloads page.

Please note that the Pulsar binary package will not contain the necessary YAML resources to deploy Pulsar on Kubernetes.

If you’d like to change the number of bookies, brokers, or ZooKeeper nodes in your Pulsar cluster, modify the replicas parameter in the spec section of the appropriate Deployment or StatefulSet resource.

Pulsar on Google Kubernetes Engine

Google Kubernetes Engine (GKE) automates the creation and management of Kubernetes clusters in Google Compute Engine (GCE).


To get started, you’ll need:

Create a new Kubernetes cluster

You can create a new GKE cluster using the container clusters create command for gcloud. This command enables you to specify the number of nodes in the cluster, the machine types of those nodes, and more.

As an example, we’ll create a new GKE cluster for Kubernetes version 1.6.4 in the us-central1-a zone. The cluster will be named pulsar-gke-cluster and will consist of three VMs, each using two locally attached SSDs and running on n1-standard-8 machines. These SSDs will be used by bookie instances, one for the BookKeeper journal and the other for storing the actual message data.

$ gcloud container clusters create pulsar-gke-cluster \
  --zone=us-central1-a \
  --machine-type=n1-standard-8 \
  --num-nodes=3 \
  --local-ssd-count=2 \

By default, bookies will run on all the machines that have locally attached SSD disks. In this example, all of those machines will have two SSDs, but you can add different types of machines to the cluster later. You can control which machines host bookie servers using labels.


You can observe your cluster in the Kubernetes Dashboard by downloading the credentials for your Kubernetes cluster and opening up a proxy to the cluster:

$ gcloud container clusters get-credentials pulsar-gke-cluster \
  --zone=us-central1-a \
$ kubectl proxy

By default, the proxy will be opened on port 8001. Now you can navigate to localhost:8001/ui in your browser to access the dashboard. At first your GKE cluster will be empty, but that will change as you begin deploying Pulsar components.

Pulsar on Amazon Web Services

You can run Kubernetes on Amazon Web Services (AWS) in a variety of ways. A very simple way that was recently introduced involves using the Kubernetes Operations (kops) tool.

You can find detailed instructions for setting up a Kubernetes cluster on AWS here.

When you create a cluster using those instructions, your kubectl config in ~/.kube/config (on MacOS and Linux) will be updated for you, so you probably won’t need to change your configuration. Nonetheless, you can ensure that kubectl can interact with your cluster by listing the nodes in the cluster:

$ kubectl get nodes

If kubectl is working with your cluster, you can proceed to deploy Pulsar components.

Pulsar on a custom Kubernetes cluster

Pulsar can be deployed on a custom, non-GKE Kubernetes cluster as well. You can find detailed documentation on how to choose a Kubernetes installation method that suits your needs in the Picking the Right Solution guide in the Kubernetes docs.

Local cluster

The easiest way to run a Kubernetes cluster is to do so locally. To install a mini local cluster for testing purposes, running in local VMs, you can either:

  1. Use minikube to run a single-node Kubernetes cluster
  2. Create a local cluster running on multiple VMs on the same machine

For the second option, follow the instructions for running Kubernetes using CoreOS on Vagrant. We’ll provide an abridged version of those instructions here.

First, make sure you have Vagrant and VirtualBox installed. Then clone the repo and start up the cluster:

$ git clone
$ cd kubernetes-vagrant-coreos-cluster

# Start a three-VM cluster
$ NODES=3 USE_KUBE_UI=true vagrant up

Create SSD disk mount points on the VMs using this script:

$ for vm in node-01 node-02 node-03; do
    NODES=3 vagrant ssh $vm -c "sudo mkdir -p /mnt/disks/ssd0"
    NODES=3 vagrant ssh $vm -c "sudo mkdir -p /mnt/disks/ssd1"

Bookies expect two logical devices to mount for journal and persistent message storage to be available. In this VM exercise, we created two directories on each VM.

Once the cluster is up, you can verify that kubectl can access it:

$ kubectl get nodes
NAME           STATUS                     AGE       VERSION   Ready,SchedulingDisabled   10m       v1.6.4   Ready                      8m        v1.6.4   Ready                      6m        v1.6.4   Ready                      4m        v1.6.4


In order to use the Kubernetes Dashboard with your local Kubernetes cluster, first use kubectl to create a proxy to the cluster:

$ kubectl proxy

Now you can access the web interface at localhost:8001/ui. At first your local cluster will be empty, but that will change as you begin deploying Pulsar components.

Deploying Pulsar components

Now that you’ve set up a Kubernetes cluster, either on Google Kubernetes Engine or on a custom cluster, you can begin deploying the components that make up Pulsar. The YAML resource definitions for Pulsar components can be found in the kubernetes folder of the Pulsar source package.

In that package, there are two sets of resource definitions, one for Google Kubernetes Engine (GKE) in the deployment/kubernetes/google-kubernetes-engine folder and one for a custom Kubernetes cluster in the deployment/kubernetes/generic folder. To begin, cd into the appropriate folder.


You must deploy ZooKeeper as the first Pulsar component, as it is a dependency for the others.

$ kubectl apply -f zookeeper.yaml

Wait until all three ZooKeeper server pods are up and have the status Running. You can check on the status of the ZooKeeper pods at any time:

$ kubectl get pods -l component=zookeeper
NAME      READY     STATUS             RESTARTS   AGE
zk-0      1/1       Running            0          18m
zk-1      1/1       Running            0          17m
zk-2      0/1       Running            6          15m

This step may take several minutes, as Kubernetes needs to download the Docker image on the VMs.

Initialize cluster metadata

Once ZooKeeper is running, you need to initialize the metadata for the Pulsar cluster in ZooKeeper. This includes system metadata for BookKeeper and Pulsar more broadly. There is a Kubernetes job in the cluster-metadata.yaml file that you only need to run once:

$ kubectl apply -f cluster-metadata.yaml

For the sake of reference, that job runs the following command on an ephemeral pod:

$ bin/pulsar initialize-cluster-metadata \
  --cluster us-central \
  --zookeeper zookeeper \
  --global-zookeeper zookeeper \
  --web-service-url http://broker.default.svc.cluster.local:8080/ \
  --broker-service-url pulsar://broker.default.svc.cluster.local:6650/

Deploy the rest of the components

Once cluster metadata has been successfully initialized, you can then deploy the bookies, brokers, monitoring stack (Prometheus, Grafana, and the Pulsar dashboard), and Pulsar cluster proxy:

$ kubectl apply -f bookie.yaml
$ kubectl apply -f broker.yaml
$ kubectl apply -f monitoring.yaml
$ kubectl apply -f proxy.yaml

You can check on the status of the pods for these components either in the Kubernetes Dashboard or using kubectl:

$ kubectl get pods -w -l app=pulsar

Set up properties and namespaces

Once all of the components are up and running, you’ll need to create at least one Pulsar property and at least one namespace.

This step is not strictly required if Pulsar authentication and authorization is turned on, though it allows you to change policies for each of the namespaces later.

You can create properties and namespaces (and perform any other administrative tasks) using the pulsar-admin pod that is already configured to act as an admin client for your newly created Pulsar cluster. One easy way to perform administrative tasks is to create an alias for the pulsar-admin tool installed on the admin pod.

$ alias pulsar-admin='kubectl exec pulsar-admin -it -- bin/pulsar-admin'

Now, any time you run pulsar-admin, you will be running commands from that pod. This command will create a property called prop:

$ pulsar-admin properties create prop \
  --admin-roles admin \
  --allowed-clusters us-central

This command will create a ns namespace under the prop property and the us-central cluster:

$ pulsar-admin namespaces create prop/us-central/ns

To verify that everything has gone as planned:

$ pulsar-admin properties list

$ pulsar-admin namespaces list prop/us-central

Now that you have a namespace and property set up, you can move on to experimenting with your Pulsar cluster from within the cluster or connecting to the cluster using a Pulsar client.

Experimenting with your cluster

Now that a property and namespace have been created, you can begin experimenting with your running Pulsar cluster. Using the same pulsar-admin pod via an alias, as in the section above, you can use pulsar-perf to create a test producer to publish 10,000 messages a second on a topic in the property and namespace you created.

First, create an alias to use the pulsar-perf tool via the admin pod:

$ alias pulsar-perf='kubectl exec pulsar-admin -it -- bin/pulsar-perf'

Now, produce messages:

$ pulsar-perf produce persistent://prop/us-central/ns/my-topic \
  --rate 10000

Similarly, you can start a consumer to subscribe to and receive all the messages on that topic:

$ pulsar-perf consume persistent://prop/us-central/ns/my-topic \
  --subscriber-name my-subscription-name

You can also view stats for the topic using the pulsar-admin tool:

$ pulsar-admin persistent stats persistent://prop/us-central/ns/my-topic


The default monitoring stack for Pulsar on Kubernetes has consists of Prometheus, Grafana, and the Pulsar dashbaord.


All Pulsar metrics in Kubernetes are collected by a Prometheus instance running inside the cluster. Typically, there is no need to access Prometheus directly. Instead, you can use the Grafana interface that displays the data stored in Prometheus.


In your Kubernetes cluster, you can use Grafana to view dashbaords for Pulsar namespaces (message rates, latency, and storage), JVM stats, ZooKeeper, and BookKeeper. You can get access to the pod serving Grafana using kubectl’s port-forward command:

$ kubectl port-forward \
  $(kubectl get pods -l component=grafana -o jsonpath='{.items[*]}') 3000

You can then access the dashboard in your web browser at localhost:3000.

Pulsar dashboard

While Grafana and Prometheus are used to provide graphs with historical data, Pulsar dashboard reports more detailed current data for individual topics.

For example, you can have sortable tables showing all namespaces, topics, and broker stats, with details on the IP address for consumers, how long they’ve been connected, and much more.

You can access to the pod serving the Pulsar dashboard using kubectl’s port-forward command:

$ kubectl port-forward \
  $(kubectl get pods -l component=dashboard -o jsonpath='{.items[*]}') 8080:80

You can then access the dashboard in your web browser at localhost:8080.

Client connections

Once your Pulsar cluster is running on Kubernetes, you can connect to it using a Pulsar client. You can fetch the IP address for the Pulsar proxy running in your Kubernetes cluster using kubectl:

$ kubectl get service broker-proxy \

If the IP address for the proxy were, for example,, you could connect to Pulsar using pulsar://

You can find client documentation for: