Join our Discord Server
Avinash Bendigeri Avinash is a developer-turned Technical writer skilled in core content creation. He has an excellent track record of blogging in areas like Docker, Kubernetes, IoT and AI.

Understanding Kubernetes Autoscaling: Vertical vs Horizontal Scaling Explained

7 min read

Say, you’ve built a fantastic application and deployed it on Kubernetes. Suddenly, a surge in traffic hits, overwhelming your cluster. Your application grinds to a halt, frustrating users. Ouch!

This is where Kubernetes Autoscaling comes in as a lifesaver. It automatically adjusts your cluster’s resources based on demand, ensuring your applications have what they need to perform optimally.

There are two scaling methods in Kubernetes : vertical scaling, and horizontal scaling. In this blog post, we will dive deep into each of these types of scaling and have a hands-on look at the way that each functions. We will also see the benefits each method has, as well as the drawbacks.

Vertical pod autoscaler

A vertical pod autoscaler works by collecting metrics (using the metrics server), and then analyzing those metrics over a period of time to understand the resource requirements of the running pods. It considers factors such as historical usage patterns, spikes in resource consumption, and the configured target utilization levels. Once this analysis is complete, the VPA controller generates recommendations for adjusting the resource requests (CPU and memory) of the pods. It may recommend increasing or decreasing resource requests to better match the observed usage. This is the basis of how a VPA works. However, this is not the end of the job for the VPA, as it has to constantly monitor and create a feedback loop where the VPA regularly adjusts pod resources based on the latest metrics.

As you might already know, these steps are also largely performed by the Horizontal pod autoscaler as well. What differentiates the VPA from the HPA is how scaling is performed. With a VPA, the autoscaler recommends changes to a pod’s resource requirements, it does so by modifying the pod’s associated resource settings in the deployment or StatefulSet manifest. This triggers Kubernetes to create new pods with the updated resource specifications and gradually replace the existing pods. So it will perform a rolling update where the old pod with insufficient resources is replaced with a new pod that has the required resource allocation.

Scaling down happens in the same way, where the VPA dynamically updates the resource specifications of existing pods. When scaling down, it may reduce the requested CPU or memory resources if historical metrics indicate that the pod consistently uses less than initially requested. Then, the VPA indirectly scales down by updating the resource settings in the pod’s associated deployment or stateful set manifest. It then triggers a controlled rolling update, creating new pods with updated resource specifications while phasing out the old ones.

Horizontal pod autoscaler

A horizontal pod autoscaler works in the same way as a VPA for the most part. It continuously monitors specified metrics, such as CPU utilization or custom metrics, for the pods it is scaling. You define a target value for the chosen metric. For example, you might set a target CPU utilization percentage. Based on the observed metrics and the defined target value, HPA makes a scaling decision to either increase or decrease the number of pod replicas. The amount of resources allocated to each pod remains the same. The number of pods will increase to accommodate this influx. If there is a service associated with the pod, the service will automatically start load balancing across the pod replicas without any intervention from your side.

Scaling down is handled in roughly the same way. When scaling down, HPA reduces the number of pod replicas. It terminates existing pods to bring the number of replicas in line with the configured target metric. The scaling decision is based on the comparison of the observed metric with the target value. HPA does not modify the resource specifications (CPU and memory requests/limits) of individual pods. Instead, it adjusts the number of replicas to match the desired metric target.

Now that we have thoroughly explored both types of autoscalers, let’s go on to a lab where we will look at the scalers in more detail.

Getting Started

You will need a Kubernetes cluster. A single node Minikube cluster will do just fine. Once the cluster is setup, you will have to install the metrics server, since the autoscalers use this to read the resource usage metrics. To do this, run:

kubectl apply -f

We will start with a base application that will have the scaling performed in it. In this case, we will use a sample nginx deployment. Create a file nginx-deployment.yaml and paste the below contents to it:

apiVersion: apps/v1
kind: Deployment
  name: nginx-deployment
  replicas: 3
      app: nginx
        app: nginx
      - name: nginx-container
        image: nginx:1.21.5
            cpu: 100m
            memory: 128Mi
apiVersion: v1
kind: Service
  name: nginx-service
    app: nginx
    - protocol: TCP
      port: 80
      targetPort: 80

This will start an nginx container that has at least 100m CPU & 128Mb memory, but not more than 200m CPU and 256Mb memory. It will also start the service that points to this deployment on port 80. Deploy this application onto your Kubernetes cluster:

kubectl apply -f nginx-deployment.yaml

Now, when the application reaches the CPU or memory limit, it will affect application performance since it is not allowed to go beyond that. So let’s introduce the autoscaler. We will start with the vertical pod autoscaler. Create a new file called “nginx-vpa.yaml” and paste the contents of the below script there.

kind: VerticalPodAutoscaler
  name: nginx-vpa
    apiVersion: "apps/v1"
    kind: Deployment
    name: nginx-deployment
    updateMode: "Auto"
    - containerName: "*"  # Apply policies to all containers in the pod
        cpu: 50m
        memory: 64Mi
        cpu: 500m
        memory: 512Mi

The resource itself is fairly self-explanatory. The spec section contains the specifications for the VPA. The targetRef section specifies the workload that the VPA is targeting for autoscaling. In this example, it’s targeting a Deployment named “nginx-deployment.” The updatePolicy section configures the update mode. In “Auto” mode, VPA automatically applies the recommended changes to the pod resources without manual intervention. The resourcePolicy section specifies the resource policies for individual containers within the pod. Within it, you have the containerPolicies section which defines policies for containers. In this case, it uses a wildcard (“*”) to apply policies to all containers in the pod. It also has the minAllowed section which specifies the minimum allowed resources. VPA won’t recommend going below these values. For example, the minimum allowed CPU is 50 milliCPU (50m), and the minimum allowed memory is 64 megabytes (64Mi). The maxAllowed section specifies the maximum allowed resources. VPA won’t recommend going above these values. For example, the maximum allowed CPU is 500 milliCPU (500m), and the maximum allowed memory is 512 megabytes (512Mi).

Now deploy this into the Kubernetes cluster:

kubectl apply -f nginx-vpa.yaml

Once the deployment is complete, we need to load-test the deployment to see the VPA in action. An important thing to note here is that if you placed the VPA memory/CPU limit too low, this will result in the pod starting up replicas immediately upon pod creation since the limit will be reached as soon as the pod comes up. This is why it is important to be aware of your average and peak loads before you begin implementing the VPA.

To load test the deployment, we will be using Apache Benchmark. Install it with apt or yum. You can do the installation on the Kubernetes node that has started. Next, note down the URL you want to load-test. To get this, use:

kubectl get svc

This will list all the services. Pick the nginx service from this list, copy its IP, and use Benchmark as below:

ab -n 1000 -c 50 http://<nginx-service-ip>/

This command will send 1000 requests with a concurrency of 50 to the NGINX service. You can adjust the -n (total requests) and -c (concurrency) parameters based on your specific load testing requirements. You can then analyze the results. Apache Benchmark will provide detailed output, including request per second (RPS), connection times, and more. For example:

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        0    1   2.8      0      10
Processing:   104  271 144.3    217    1184
Waiting:      104  270 144.2    217    1184
Total:        104  272 144.5    217    1185

Now it’s time to check if autoscaling has started:

kubectl get po -n default

Watch the pods, and you will see that the resource limits are reached, after which a new pod with more resources is created. Keep an eye on the resource usage and you will notice that the new resources have higher limits. Once the requests have been handled, the pod will immediately reduce the resource consumption. However, a new pod with lower resource requirements will not show up to replace the old pod. In fact, if you were to push a new version of the deployment into the cluster, it would still have space for a large amount of requests. However, this will reduce eventually if the amount of resources consumed continues to be low.

Now that we have gotten a complete look at the vertical pod autoscaler, let’s take a look at the HPA. Create a file nginx-hpa.yml and paste the below contents into it.

apiVersion: autoscaling/v2beta2
kind: HorizontalPodAutoscaler
  name: nginx-hpa
    apiVersion: apps/v1
    kind: Deployment
    name: nginx-deployment
  minReplicas: 1
  maxReplicas: 5
  - type: Resource
      name: cpu
        type: Utilization
        averageUtilization: 80

The above HPA definition has a lot of similarities to the VPA definition. The differences lie in the minReplicas and maxReplicas sections which define the minimum and maximum number of pod replicas that the HPA should maintain. In this case, it’s set to have a minimum of 2 replicas and a maximum of 5 replicas. The VPA didn’t have a metrics section that the HPA has, but its resourcePolicy section is pretty similar to this, where the metrics configure the metric used for autoscaling. In this example, it’s using the CPU utilization metric.type: Resource: Specifies that the metric is a resource metric (in this case, CPU). The resource section specifies the resource metric details. name: cpu Indicates that the metric is CPU utilization. The target section specifies the target value for the metric and type: Utilization indicates that the target is based on resource utilization. averageUtilization sets the target average CPU utilization to 80%.

Before you deploy this file into your cluster, make sure to remove the VPA since having two types of autoscalers running for the same pod can cause some obvious problems. So first run:

kubectl delete -f nginx-vpa.yaml

Then deploy the HPA:

kubectl apply -f nginx-hpa.yaml

You can see the status of the HPA as it starts up using describe:

kubectl describe hpa nginx-hpa

You might see some errors about the HPA being unable to retrieve metrics, however, these can be ignored since this is an issue that occurs only when the HPA starts up for the first time. Now, let’s go back to the apache benchmark and add load to the nginx service so that we can see the HPA in action. Let’s start it up in the same manner as before:

ab -n 1000 -c 50 http://<nginx-service-ip>/

A thousand requests should start being sent to the service. Start watching the nginx pod to see if replicas are being created:

kubectl get po -n default --watch

You should be able to see the memory limit getting reached, after which the number of pods will increase. This will keep happening until the number of pods reaches the maximum specified value (5) or the memory requests are satisfied.


That sums up the lab on autoscalers. In here, we discussed the two most commonly used in-built autoscalers: HPA and VPA. We also took a hands-on look at how the autoscalers worked. This is just the tip of the iceberg when it comes to scaling, however, and the subject of custom scalers that can scale based on metrics other than memory and CPU is vast. If you are interested in looking at more complicated scaling techniques, you could take a look at the KEDA section to get some idea of the keda autoscaler.

Have Queries? Join

Avinash Bendigeri Avinash is a developer-turned Technical writer skilled in core content creation. He has an excellent track record of blogging in areas like Docker, Kubernetes, IoT and AI.
Join our Discord Server