How to create a Kubernetes pod disruption budget using Gremlin

How to create a Kubernetes pod disruption budget using Gremlin
Last Updated:
Categories: Chaos Engineering


In this tutorial, we'll show you how to create and validate a pod disruption budget (PDB) for a Kubernetes application. We'll walk through creating a PDB for a replicated Deployment, deploying the Gremlin Kubernetes agent, running a Shutdown attack on a node, then validating that Kubernetes enforces our PDB.

By reading this tutorial, you'll learn how to create a PDB and validate that it works as expected by running a Shutdown attack.


This tutorial will show you how to:

  • Step 1 - Deploy a multi-node Kubernetes cluster
  • Step 2 - Deploy the Gremlin Kubernetes agent
  • Step 3 - Deploy an application with a Pod Disruption Budget
  • Step 4 - Run a Shutdown experiment to validate your Pod Disruption Budget


In Kubernetes lingo, anything that could potentially interrupt the operation of a Pod is called a disruption. This can be anything from a system crash causing a node to reboot, to an SRE accidentally typing kubectl delete instead of kubectl apply . If a disruption is caused by an operator—such as deleting a Deployment or draining a node—it's known as a voluntary disruption.

A pod disruption budget (PDB) lets application owners set requirements around Deployments so that they're less likely to be impacted by voluntary disruptions. In essence, it lets you define the minimum tolerable operational requirements for a Deployment so that it remains stable even during a disruption.

For example, imagine we have a Deployment that contains four replicas distributed evenly across four nodes. We can tolerate losing one Pod, but not two, so we create a PDB that requires a minimum of three replicas:

1apiVersion: policy/v1
2kind: PodDisruptionBudget
4 name: my-app-pdb
6 minAvailable: 3
7 selector:
8 matchLabels:
9 app: my-app

We want to scale down our cluster to save costs, so we reduce our node count to two. However, this would also reduce the size of the my-app Deployment to two replicas. Kubernetes recognizes this, and because of the PDB, creates a new instance of the Deployment on one of the remaining nodes. This ensures that there are at least three replicas and allows the nodes to terminate.


Before starting this tutorial, you'll need the following:

  • A Gremlin account
  • A Kubernetes cluster with at least two worker nodes
  • The kubectl command-line tool
  • The Helm command-line tool

Step 1 - Deploy a multi-node Kubernetes cluster

For this tutorial, you'll need a Kubernetes cluster with at least two worker nodes. If you already have a cluster, you can continue to step 2. If not, you can create your own cluster using a tool like kubeadm or K3s, or use a cloud Kubernetes platform like Amazon EKS, Azure Kubernetes Service, or Google Kubernetes Engine.

Once your cluster is deployed, verify that you have at least two actively running worker nodes by using kubectl get nodes :

2gremlin-lab-control-plane Ready control-plane,master 7d v1.21.1
3gremlin-lab-worker Ready <none> 7d v1.21.1
4gremlin-lab-worker2 Ready <none> 7d v1.21.1

Step 2 - Deploy the Gremlin Kubernetes agent

The Gremlin Kubernetes agent is the recommended method of deploying Gremlin to a Kubernetes cluster. It allows Gremlin to detect resources such as Pods, Deployments, and DaemonSets, and select them as targets for running experiments.

To install the agent, follow the installation instructions in the documentation. Make sure to replace $GREMLIN_TEAM_ID with your actual team ID, $GREMLIN_TEAM_SECRET with your team secret key, and $GREMLIN_CLUSTER_ID with a uniquely identifying name for this cluster. You can learn more about team IDs and secret keys in the Authentication documentation.

Once the agent has been installed, you can confirm that it's working correctly by logging into your Gremlin account, opening the Clients page, then clicking the Kubernetes tab. Your cluster will be listed with the name provided by $GREMLIN_CLUSTER_ID :

Viewing a list of Kubernetes clusters in the Gremlin web app

Step 3 - Deploy an application with a Pod Disruption Budget

Next, let's deploy an application and create a Pod Disruption Budget. For this example, we'll use the open source Online Boutique demo application created by the Google Cloud Platform team. Online Boutique is an open source e-commerce website consisting of 12 services, each one in its own Deployment. To deploy the application, either read the quick start guide or simply run the following command:

1kubectl apply -f

You can check on the status of the deployment by running:

1kubectl get deployments
2default adservice 1/1 1 1 5d21h
3default cartservice 1/1 1 1 5d21h
4default checkoutservice 1/1 1 1 5d21h
5default currencyservice 1/1 1 1 5d21h
6default emailservice 1/1 1 1 5d21h
7default frontend 1/1 1 1 5d21h
8default loadgenerator 1/1 1 1 5d21h
9default paymentservice 1/1 1 1 5d21h
10default productcatalogservice 1/1 1 1 5d21h
11default recommendationservice 1/1 1 1 5d21h
12default redis-cart 1/1 1 1 5d21h
13default shippingservice 1/1 1 1 5d21h

By default, each microservice only deploys a single replica. We're going to pick a service to create a PDB for: for this tutorial, let's use the frontend service and increase the replica count to 2:

1kubectl scale --replicas=2 deployments/frontend
1deployment.apps/frontend scaled

Next, let's create our PDB. Our PDB will ensure that there are always at least 2 replicas running at any given time. This should prevent us from shutting down one of our nodes if the number of remaining Pods is less than 2:

1apiVersion: policy/v1
2kind: PodDisruptionBudget
4 name: frontend-pdb
6 minAvailable: 2
7 selector:
8 matchLabels:
9 app: frontend

Save this file as frontend-pdb.yaml, then apply it to your cluster:

1kubectl apply -f frontend-pdb.yaml

You'll see the following output:

1poddisruptionbudget.policy/frontend-pdb created

Step 4 - Run a Shutdown experiment to validate your Pod Disruption Budget

The final step is to ensure your PDB works by running an experiment. We'll use a Shutdown attack on one of our worker nodes and monitor both the node and our frontend Deployment to see whether our PDB is honored.

First, log into the Gremlin web app and create a new attack by navigating to Attacks, then New Attack. Select the Infrastructure tab, then selects Hosts (if not already selected). Select one of your worker nodes, or select all of your worker nodes and change the Percent of targets to impact to 50 (you could also enter 1 and use the drop-down to select # instead of %).

Next, click on Choose a Gremlin to expand the attack selection box, select the State category, then select Shutdown. Reduce the Delay to 0 to immediately start the attack, and make sure Reboot is checked. Finally, click Unleash Gremlin to run the attack.

While the attack is starting, open a terminal and retrieve the state of your PDB by using the command kubectl get poddisruptionbudgets frontend-pdb -o yaml. This should return something similar to the following:

1apiVersion: policy/v1
2kind: PodDisruptionBudget
4 annotations:
6 creationTimestamp: "2022-01-04T04:22:56Z"
7 generation: 1
8 name: frontend-pdb
11 currentHealthy: 2
12 desiredHealthy: 2
13 disruptionsAllowed: 0
14 expectedPods: 2
15 observedGeneration: 1

Note the currentHealthy and disruptionsAllowed fields, as these will tell us the number of actively running Pods and the maximum number of failed Pods allowed respectively.

Meanwhile, we should also monitor the status of our nodes by using kubectl or an external monitoring tool:

1watch - n 2 'kubectl get nodes'

What we should see is currentHealthy fall no lower than 2, even if it means one of our workers takes longer to shutdown. Ideally it should remain at 3 as Kubernetes reschedules replicas onto the remaining worker node. If it goes down to 1, then we can assume both workers shutdown despite our settings, and we should therefore recheck our PDB configuration or reevaluate our application deployment strategy.


In this tutorial, we used Gremlin to ensure that our deployments can withstand voluntary disruptions in our Kubernetes cluster. Our applications are ready to autoscale, deploy onto Spot instances, and handle even large-scale cluster disruptions. To learn more about Shutdown attacks, read our blog on Getting started with Shutdown attacks.


Avoid downtime. Use Gremlin to turn failure into resilience.

Gremlin empowers you to proactively root out failure before it causes downtime. See how you can harness chaos to build resilient systems by requesting a demo of Gremlin.

Get started