1. Documentation
    2. history bug_report picture_as_pdf
×

Automatically adjust pod resource levels with the vertical pod autoscaler

The OpenShift Container Platform Vertical Pod Autoscaler Operator (VPA) automatically reviews the historic and current CPU and memory resources for containers in pods and can update the resource limits and requests based on the usage values it learns. The VPA uses individual custom resources (CR) to update all of the pods associated with a workload object, such as a Deployment, DeploymentConfig, StatefulSet, Job, DaemonSet, ReplicaSet, or ReplicationController, in a project.

The VPA helps you to understand the optimal CPU and memory usage for your pods and can automatically maintain pod resources through the pod lifecycle.

About the Vertical Pod Autoscaler Operator

The Vertical Pod Autoscaler Operator (VPA) is implemented as an API resource and a custom resource (CR). The CR determines the actions the Vertical Pod Autoscaler Operator should take with the pods associated with a specific workload object, such as a daemon set, replication controller, and so forth, in a project.

You can use the default recommender or use your own alternative recommender to autoscale based on your own algorithms.

The default recommender automatically computes historic and current CPU and memory usage for the containers in those pods and uses this data to determine optimized resource limits and requests to ensure that these pods are operating efficiently at all times. For example, the default recommender suggests reduced resources for pods that are requesting more resources than they are using and increased resources for pods that are not requesting enough.

The VPA then automatically deletes any pods that are out of alignment with these recommendations one at a time, so that your applications can continue to serve requests with no downtime. The workload objects then re-deploy the pods with the original resource limits and requests. The VPA uses a mutating admission webhook to update the pods with optimized resource limits and requests before the pods are admitted to a node. If you do not want the VPA to delete pods, you can view the VPA resource limits and requests and manually update the pods as needed.

By default, workload objects must specify a minimum of two replicas in order for the VPA to automatically delete their pods. Workload objects that specify fewer replicas than this minimum are not deleted. If you manually delete these pods, when the workload object redeploys the pods, the VPA does update the new pods with its recommendations. You can change this minimum by modifying the VerticalPodAutoscalerController object as shown shown in Changing the VPA minimum value.

For example, if you have a pod that uses 50% of the CPU but only requests 10%, the VPA determines that the pod is consuming more CPU than requested and deletes the pod. The workload object, such as replica set, restarts the pods and the VPA updates the new pod with its recommended resources.

For developers, you can use the VPA to help ensure your pods stay up during periods of high demand by scheduling pods onto nodes that have appropriate resources for each pod.

Administrators can use the VPA to better utilize cluster resources, such as preventing pods from reserving more CPU resources than needed. The VPA monitors the resources that workloads are actually using and adjusts the resource requirements so capacity is available to other workloads. The VPA also maintains the ratios between limits and requests that are specified in initial container configuration.

If you stop running the VPA or delete a specific VPA CR in your cluster, the resource requests for the pods already modified by the VPA do not change. Any new pods get the resources defined in the workload object, not the previous recommendations made by the VPA.

Installing the Vertical Pod Autoscaler Operator

You can use the OpenShift Container Platform web console to install the Vertical Pod Autoscaler Operator (VPA).

Procedure

  1. In the OpenShift Container Platform web console, click Operators → OperatorHub.

  2. Choose VerticalPodAutoscaler from the list of available Operators, and click Install.

  3. On the Install Operator page, ensure that the Operator recommended namespace option is selected. This installs the Operator in the mandatory openshift-vertical-pod-autoscaler namespace, which is automatically created if it does not exist.

  4. Click Install.

  5. Verify the installation by listing the VPA Operator components:

    1. Navigate to Workloads → Pods.

    2. Select the openshift-vertical-pod-autoscaler project from the drop-down menu and verify that there are four pods running.

    3. Navigate to Workloads → Deployments to verify that there are four deployments running.

  6. Optional. Verify the installation in the OpenShift Container Platform CLI using the following command:

    $ oc get all -n openshift-vertical-pod-autoscaler

    The output shows four pods and four deplyoments:

    Example output
    NAME                                                    READY   STATUS    RESTARTS   AGE
pod/vertical-pod-autoscaler-operator-85b4569c47-2gmhc   1/1     Running   0          3m13s
pod/vpa-admission-plugin-default-67644fc87f-xq7k9       1/1     Running   0          2m56s
pod/vpa-recommender-default-7c54764b59-8gckt            1/1     Running   0          2m56s
pod/vpa-updater-default-7f6cc87858-47vw9                1/1     Running   0          2m56s

NAME                  TYPE        CLUSTER-IP      EXTERNAL-IP   PORT(S)   AGE
service/vpa-webhook   ClusterIP   172.30.53.206   <none>        443/TCP   2m56s

NAME                                               READY   UP-TO-DATE   AVAILABLE   AGE
deployment.apps/vertical-pod-autoscaler-operator   1/1     1            1           3m13s
deployment.apps/vpa-admission-plugin-default       1/1     1            1           2m56s
deployment.apps/vpa-recommender-default            1/1     1            1           2m56s
deployment.apps/vpa-updater-default                1/1     1            1           2m56s

NAME                                                          DESIRED   CURRENT   READY   AGE
replicaset.apps/vertical-pod-autoscaler-operator-85b4569c47   1         1         1       3m13s
replicaset.apps/vpa-admission-plugin-default-67644fc87f       1         1         1       2m56s
replicaset.apps/vpa-recommender-default-7c54764b59            1         1         1       2m56s
replicaset.apps/vpa-updater-default-7f6cc87858                1         1         1       2m56s

About Using the Vertical Pod Autoscaler Operator

To use the Vertical Pod Autoscaler Operator (VPA), you create a VPA custom resource (CR) for a workload object in your cluster. The VPA learns and applies the optimal CPU and memory resources for the pods associated with that workload object. You can use a VPA with a deployment, stateful set, job, daemon set, replica set, or replication controller workload object. The VPA CR must be in the same project as the pods you want to monitor.

You use the VPA CR to associate a workload object and specify which mode the VPA operates in:

  • The Auto and Recreate modes automatically apply the VPA CPU and memory recommendations throughout the pod lifetime. The VPA deletes any pods in the project that are out of alignment with its recommendations. When redeployed by the workload object, the VPA updates the new pods with its recommendations.

  • The Initial mode automatically applies VPA recommendations only at pod creation.

  • The Off mode only provides recommended resource limits and requests, allowing you to manually apply the recommendations. The off mode does not update pods.

You can also use the CR to opt-out certain containers from VPA evaluation and updates.

For example, a pod has the following limits and requests:

resources:
  limits:
    cpu: 1
    memory: 500Mi
  requests:
    cpu: 500m
    memory: 100Mi

After creating a VPA that is set to auto, the VPA learns the resource usage and deletes the pod. When redeployed, the pod uses the new resource limits and requests:

resources:
  limits:
    cpu: 50m
    memory: 1250Mi
  requests:
    cpu: 25m
    memory: 262144k

You can view the VPA recommendations using the following command:

$ oc get vpa <vpa-name> --output yaml

After a few minutes, the output shows the recommendations for CPU and memory requests, similar to the following:

Example output
...
status:
...
  recommendation:
    containerRecommendations:
    - containerName: frontend
      lowerBound:
        cpu: 25m
        memory: 262144k
      target:
        cpu: 25m
        memory: 262144k
      uncappedTarget:
        cpu: 25m
        memory: 262144k
      upperBound:
        cpu: 262m
        memory: "274357142"
    - containerName: backend
      lowerBound:
        cpu: 12m
        memory: 131072k
      target:
        cpu: 12m
        memory: 131072k
      uncappedTarget:
        cpu: 12m
        memory: 131072k
      upperBound:
        cpu: 476m
        memory: "498558823"
...

The output shows the recommended resources, target, the minimum recommended resources, lowerBound, the highest recommended resources, upperBound, and the most recent resource recommendations, uncappedTarget.

The VPA uses the lowerBound and upperBound values to determine if a pod needs to be updated. If a pod has resource requests below the lowerBound values or above the upperBound values, the VPA terminates and recreates the pod with the target values.

Changing the VPA minimum value

By default, workload objects must specify a minimum of two replicas in order for the VPA to automatically delete and update their pods. As a result, workload objects that specify fewer than two replicas are not automatically acted upon by the VPA. The VPA does update new pods from these workload objects if the pods are restarted by some process external to the VPA. You can change this cluster-wide minimum value by modifying the minReplicas parameter in the VerticalPodAutoscalerController custom resource (CR).

For example, if you set minReplicas to 3, the VPA does not delete and update pods for workload objects that specify fewer than three replicas.

If you set minReplicas to 1, the VPA can delete the only pod for a workload object that specifies only one replica. You should use this setting with one-replica objects only if your workload can tolerate downtime whenever the VPA deletes a pod to adjust its resources. To avoid unwanted downtime with one-replica objects, configure the VPA CRs with the podUpdatePolicy set to Initial, which automatically updates the pod only when it is restarted by some process external to the VPA, or Off, which allows you to update the pod manually at an appropriate time for your application.
Example VerticalPodAutoscalerController object
apiVersion: autoscaling.openshift.io/v1
kind: VerticalPodAutoscalerController
metadata:
  creationTimestamp: "2021-04-21T19:29:49Z"
  generation: 2
  name: default
  namespace: openshift-vertical-pod-autoscaler
  resourceVersion: "142172"
  uid: 180e17e9-03cc-427f-9955-3b4d7aeb2d59
spec:
  minReplicas: 3 (1)
  podMinCPUMillicores: 25
  podMinMemoryMb: 250
  recommendationOnly: false
  safetyMarginFraction: 0.15
1 Specify the minimum number of replicas in a workload object for the VPA to act on. Any objects with replicas fewer than the minimum are not automatically deleted by the VPA.

Automatically applying VPA recommendations

To use the VPA to automatically update pods, create a VPA CR for a specific workload object with updateMode set to Auto or Recreate.

When the pods are created for the workload object, the VPA constantly monitors the containers to analyze their CPU and memory needs. The VPA deletes any pods that do not meet the VPA recommendations for CPU and memory. When redeployed, the pods use the new resource limits and requests based on the VPA recommendations, honoring any pod disruption budget set for your applications. The recommendations are added to the status field of the VPA CR for reference.

By default, workload objects must specify a minimum of two replicas in order for the VPA to automatically delete their pods. Workload objects that specify fewer replicas than this minimum are not deleted. If you manually delete these pods, when the workload object redeploys the pods, the VPA does update the new pods with its recommendations. You can change this minimum by modifying the VerticalPodAutoscalerController object as shown shown in Changing the VPA minimum value.
Example VPA CR for the Auto mode
apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: vpa-recommender
spec:
  targetRef:
    apiVersion: "apps/v1"
    kind:       Deployment (1)
    name:       frontend (2)
  updatePolicy:
    updateMode: "Auto" (3)
1 The type of workload object you want this VPA CR to manage.
2 The name of the workload object you want this VPA CR to manage.
3 Set the mode to Auto or Recreate:

  • Auto. The VPA assigns resource requests on pod creation and updates the existing pods by terminating them when the requested resources differ significantly from the new recommendation.

  • Recreate. The VPA assigns resource requests on pod creation and updates the existing pods by terminating them when the requested resources differ significantly from the new recommendation. This mode should be used rarely, only if you need to ensure that the pods are restarted whenever the resource request changes.

There must be operating pods in the project before the VPA can determine recommended resources and apply the recommendations to new pods.

Automatically applying VPA recommendations on pod creation

To use the VPA to apply the recommended resources only when a pod is first deployed, create a VPA CR for a specific workload object with updateMode set to Initial.

Then, manually delete any pods associated with the workload object that you want to use the VPA recommendations. In the Initial mode, the VPA does not delete pods and does not update the pods as it learns new resource recommendations.

Example VPA CR for the Initial mode
apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: vpa-recommender
spec:
  targetRef:
    apiVersion: "apps/v1"
    kind:       Deployment (1)
    name:       frontend (2)
  updatePolicy:
    updateMode: "Initial" (3)
1 The type of workload object you want this VPA CR to manage.
2 The name of the workload object you want this VPA CR to manage.
3 Set the mode to Initial. The VPA assigns resources when pods are created and does not change the resources during the lifetime of the pod.

There must be operating pods in the project before a VPA can determine recommended resources and apply the recommendations to new pods.

Manually applying VPA recommendations

To use the VPA to only determine the recommended CPU and memory values, create a VPA CR for a specific workload object with updateMode set to off.

When the pods are created for that workload object, the VPA analyzes the CPU and memory needs of the containers and records those recommendations in the status field of the VPA CR. The VPA does not update the pods as it determines new resource recommendations.

Example VPA CR for the Off mode
apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: vpa-recommender
spec:
  targetRef:
    apiVersion: "apps/v1"
    kind:       Deployment (1)
    name:       frontend (2)
  updatePolicy:
    updateMode: "Off" (3)
1 The type of workload object you want this VPA CR to manage.
2 The name of the workload object you want this VPA CR to manage.
3 Set the mode to Off.

You can view the recommendations using the following command.

$ oc get vpa <vpa-name> --output yaml

With the recommendations, you can edit the workload object to add CPU and memory requests, then delete and redeploy the pods using the recommended resources.

There must be operating pods in the project before a VPA can determine recommended resources.

Exempting containers from applying VPA recommendations

If your workload object has multiple containers and you do not want the VPA to evaluate and act on all of the containers, create a VPA CR for a specific workload object and add a resourcePolicy to opt-out specific containers.

When the VPA updates the pods with recommended resources, any containers with a resourcePolicy are not updated and the VPA does not present recommendations for those containers in the pod.

apiVersion: autoscaling.k8s.io/v1
kind: VerticalPodAutoscaler
metadata:
  name: vpa-recommender
spec:
  targetRef:
    apiVersion: "apps/v1"
    kind:       Deployment (1)
    name:       frontend (2)
  updatePolicy:
    updateMode: "Auto" (3)
  resourcePolicy: (4)
    containerPolicies:
    - containerName: my-opt-sidecar
      mode: "Off"
1 The type of workload object you want this VPA CR to manage.
2 The name of the workload object you want this VPA CR to manage.
3 Set the mode to Auto, Recreate, or Off. The Recreate mode should be used rarely, only if you need to ensure that the pods are restarted whenever the resource request changes.
4 Specify the containers you want to opt-out and set mode to Off.

For example, a pod has two containers, the same resource requests and limits:

# ...
spec:
  containers:
  - name: frontend
    resources:
      limits:
        cpu: 1
        memory: 500Mi
      requests:
        cpu: 500m
        memory: 100Mi
  - name: backend
    resources:
      limits:
        cpu: "1"
        memory: 500Mi
      requests:
        cpu: 500m
        memory: 100Mi
# ...

After launching a VPA CR with the backend container set to opt-out, the VPA terminates and recreates the pod with the recommended resources applied only to the frontend container:

...
spec:
  containers:
    name: frontend
    resources:
      limits:
        cpu: 50m
        memory: 1250Mi
      requests:
        cpu: 25m
        memory: 262144k
...
    name: backend
    resources:
      limits:
        cpu: "1"
        memory: 500Mi
      requests:
        cpu: 500m
        memory: 100Mi
...

Using an alternative recommender

You can use your own recommender to autoscale based on your own algorithms. If you do not specify an alternative recommender, OpenShift Container Platform uses the default recommender, which suggests CPU and memory requests based on historical usage. Because there is no universal recommendation policy that applies to all types of workloads, you might want to create and deploy different recommenders for specific workloads.

For example, the default recommender might not accurately predict future resource usage when containers exhibit certain resource behaviors, such as cyclical patterns that alternate between usage spikes and idling as used by monitoring applications, or recurring and repeating patterns used with deep learning applications. Using the default recommender with these usage behaviors might result in significant over-provisioning and Out of Memory (OOM) kills for your applications.