As a developer, you can use a horizontal pod autoscaler (HPA) to specify how OKD should automatically increase or decrease the scale of a replication controller or deployment configuration, based on metrics collected from the pods that belong to that replication controller or deployment configuration. You can create an HPA for any any deployment, deployment config, replica set, replication controller, or stateful set.

For information on scaling pods based on custom metrics, see Automatically scaling pods based on custom metrics.

It is recommended to use a Deployment object or ReplicaSet object unless you need a specific feature or behavior provided by other objects. For more information on these objects, see Understanding Deployment and DeploymentConfig objects.

Understanding horizontal pod autoscalers

You can create a horizontal pod autoscaler to specify the minimum and maximum number of pods you want to run, as well as the CPU utilization or memory utilization your pods should target.

After you create a horizontal pod autoscaler, OKD begins to query the CPU and/or memory resource metrics on the pods. When these metrics are available, the horizontal pod autoscaler computes the ratio of the current metric utilization with the desired metric utilization, and scales up or down accordingly. The query and scaling occurs at a regular interval, but can take one to two minutes before metrics become available.

For replication controllers, this scaling corresponds directly to the replicas of the replication controller. For deployment configurations, scaling corresponds directly to the replica count of the deployment configuration. Note that autoscaling applies only to the latest deployment in the Complete phase.

OKD automatically accounts for resources and prevents unnecessary autoscaling during resource spikes, such as during start up. Pods in the unready state have 0 CPU usage when scaling up and the autoscaler ignores the pods when scaling down. Pods without known metrics have 0% CPU usage when scaling up and 100% CPU when scaling down. This allows for more stability during the HPA decision. To use this feature, you must configure readiness checks to determine if a new pod is ready for use.

To use horizontal pod autoscalers, your cluster administrator must have properly configured cluster metrics.

Supported metrics

The following metrics are supported by horizontal pod autoscalers:

Table 1. Metrics
Metric Description API version

CPU utilization

Number of CPU cores used. Can be used to calculate a percentage of the pod’s requested CPU.

autoscaling/v1, autoscaling/v2

Memory utilization

Amount of memory used. Can be used to calculate a percentage of the pod’s requested memory.


For memory-based autoscaling, memory usage must increase and decrease proportionally to the replica count. On average:

  • An increase in replica count must lead to an overall decrease in memory (working set) usage per-pod.

  • A decrease in replica count must lead to an overall increase in per-pod memory usage.

Use the OKD web console to check the memory behavior of your application and ensure that your application meets these requirements before using memory-based autoscaling.

The following example shows autoscaling for the image-registry Deployment object. The initial deployment requires 3 pods. The HPA object increases the minimum to 5. If CPU usage on the pods reaches 75%, the pods increase to 7:

$ oc autoscale deployment/image-registry --min=5 --max=7 --cpu-percent=75
Example output
horizontalpodautoscaler.autoscaling/image-registry autoscaled
Sample HPA for the image-registry Deployment object with minReplicas set to 3
apiVersion: autoscaling/v1
kind: HorizontalPodAutoscaler
  name: image-registry
  namespace: default
  maxReplicas: 7
  minReplicas: 3
    apiVersion: apps/v1
    kind: Deployment
    name: image-registry
  targetCPUUtilizationPercentage: 75
  currentReplicas: 5
  desiredReplicas: 0
  1. View the new state of the deployment:

    $ oc get deployment image-registry

    There are now 5 pods in the deployment:

    Example output
    image-registry   1          5         5         config

How does the HPA work?

The horizontal pod autoscaler (HPA) extends the concept of pod auto-scaling. The HPA lets you create and manage a group of load-balanced nodes. The HPA automatically increases or decreases the number of pods when a given CPU or memory threshold is crossed.

Figure 1. High level workflow of the HPA

The HPA is an API resource in the Kubernetes autoscaling API group. The autoscaler works as a control loop with a default of 15 seconds for the sync period. During this period, the controller manager queries the CPU, memory utilization, or both, against what is defined in the YAML file for the HPA. The controller manager obtains the utilization metrics from the resource metrics API for per-pod resource metrics like CPU or memory, for each pod that is targeted by the HPA.

If a utilization value target is set, the controller calculates the utilization value as a percentage of the equivalent resource request on the containers in each pod. The controller then takes the average of utilization across all targeted pods and produces a ratio that is used to scale the number of desired replicas. The HPA is configured to fetch metrics from metrics.k8s.io, which is provided by the metrics server. Because of the dynamic nature of metrics evaluation, the number of replicas can fluctuate during scaling for a group of replicas.

To implement the HPA, all targeted pods must have a resource request set on their containers.

About requests and limits

The scheduler uses the resource request that you specify for containers in a pod, to decide which node to place the pod on. The kubelet enforces the resource limit that you specify for a container to ensure that the container is not allowed to use more than the specified limit. The kubelet also reserves the request amount of that system resource specifically for that container to use.

How to use resource metrics?

In the pod specifications, you must specify the resource requests, such as CPU and memory. The HPA uses this specification to determine the resource utilization and then scales the target up or down.

For example, the HPA object uses the following metric source:

type: Resource
  name: cpu
    type: Utilization
    averageUtilization: 60

In this example, the HPA keeps the average utilization of the pods in the scaling target at 60%. Utilization is the ratio between the current resource usage to the requested resource of the pod.

Best practices

All pods must have resource requests configured

The HPA makes a scaling decision based on the observed CPU or memory utilization values of pods in an OKD cluster. Utilization values are calculated as a percentage of the resource requests of each pod. Missing resource request values can affect the optimal performance of the HPA.

Configure the cool down period

During horizontal pod autoscaling, there might be a rapid scaling of events without a time gap. Configure the cool down period to prevent frequent replica fluctuations. You can specify a cool down period by configuring the stabilizationWindowSeconds field. The stabilization window is used to restrict the fluctuation of replicas count when the metrics used for scaling keep fluctuating. The autoscaling algorithm uses this window to infer a previous desired state and avoid unwanted changes to workload scale.

For example, a stabilization window is specified for the scaleDown field:

    stabilizationWindowSeconds: 300

In the above example, all desired states for the past 5 minutes are considered. This approximates a rolling maximum, and avoids having the scaling algorithm frequently remove pods only to trigger recreating an equivalent pod just moments later.

Scaling policies

The autoscaling/v2 API allows you to add scaling policies to a horizontal pod autoscaler. A scaling policy controls how the OKD horizontal pod autoscaler (HPA) scales pods. Scaling policies allow you to restrict the rate that HPAs scale pods up or down by setting a specific number or specific percentage to scale in a specified period of time. You can also define a stabilization window, which uses previously computed desired states to control scaling if the metrics are fluctuating. You can create multiple policies for the same scaling direction, and determine which policy is used, based on the amount of change. You can also restrict the scaling by timed iterations. The HPA scales pods during an iteration, then performs scaling, as needed, in further iterations.

Sample HPA object with a scaling policy
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
  name: hpa-resource-metrics-memory
  namespace: default
    scaleDown: (1)
      policies: (2)
      - type: Pods (3)
        value: 4 (4)
        periodSeconds: 60 (5)
      - type: Percent
        value: 10 (6)
        periodSeconds: 60
      selectPolicy: Min (7)
      stabilizationWindowSeconds: 300 (8)
    scaleUp: (9)
      - type: Pods
        value: 5 (10)
        periodSeconds: 70
      - type: Percent
        value: 12 (11)
        periodSeconds: 80
      selectPolicy: Max
      stabilizationWindowSeconds: 0
1 Specifies the direction for the scaling policy, either scaleDown or scaleUp. This example creates a policy for scaling down.
2 Defines the scaling policy.
3 Determines if the policy scales by a specific number of pods or a percentage of pods during each iteration. The default value is pods.
4 Determines the amount of scaling, either the number of pods or percentage of pods, during each iteration. There is no default value for scaling down by number of pods.
5 Determines the length of a scaling iteration. The default value is 15 seconds.
6 The default value for scaling down by percentage is 100%.
7 Determines which policy to use first, if multiple policies are defined. Specify Max to use the policy that allows the highest amount of change, Min to use the policy that allows the lowest amount of change, or Disabled to prevent the HPA from scaling in that policy direction. The default value is Max.
8 Determines the time period the HPA should look back at desired states. The default value is 0.
9 This example creates a policy for scaling up.
10 The amount of scaling up by the number of pods. The default value for scaling up the number of pods is 4%.
11 The amount of scaling up by the percentage of pods. The default value for scaling up by percentage is 100%.
Example policy for scaling down
apiVersion: autoscaling/v2
kind: HorizontalPodAutoscaler
  name: hpa-resource-metrics-memory
  namespace: default
  minReplicas: 20
      stabilizationWindowSeconds: 300
      - type: Pods
        value: 4
        periodSeconds: 30
      - type: Percent
        value: 10
        periodSeconds: 60
      selectPolicy: Max
      selectPolicy: Disabled

In this example, when the number of pods is greater than 40, the percent-based policy is used for scaling down, as that policy results in a larger change, as required by the selectPolicy.

If there are 80 pod replicas, in the first iteration the HPA reduces the pods by 8, which is 10% of the 80 pods (based on the type: Percent and value: 10 parameters), over one minute (periodSeconds: 60). For the next iteration, the number of pods is 72. The HPA calculates that 10% of the remaining pods is 7.2, which it rounds up to 8 and scales down 8 pods. On each subsequent iteration, the number of pods to be scaled is re-calculated based on the number of remaining pods. When the number of pods falls below 40, the pods-based policy is applied, because the pod-based number is greater than the percent-based number. The HPA reduces 4 pods at a time (type: Pods and value: 4), over 30 seconds (periodSeconds: 30), until there are 20 replicas remaining (minReplicas).

The selectPolicy: Disabled parameter prevents the HPA from scaling up the pods. You can manually scale up by adjusting the number of replicas in the replica set or deployment set, if needed.

If set, you can view the scaling policy by using the oc edit command:

$ oc edit hpa hpa-resource-metrics-memory
Example output
apiVersion: autoscaling/v1
kind: HorizontalPodAutoscaler

Creating a horizontal pod autoscaler by using the web console

From the web console, you can create a horizontal pod autoscaler (HPA) that specifies the minimum and maximum number of pods you want to run on a Deployment or DeploymentConfig object. You can also define the amount of CPU or memory usage that your pods should target.

An HPA cannot be added to deployments that are part of an Operator-backed service, Knative service, or Helm chart.


To create an HPA in the web console:

  1. In the Topology view, click the node to reveal the side pane.

  2. From the Actions drop-down list, select Add HorizontalPodAutoscaler to open the Add HorizontalPodAutoscaler form.