Tasks

Kubernetes v1.16 documentation is no longer actively maintained. The version you are currently viewing is a static snapshot. For up-to-date documentation, see the latest version.

Edit This Page

Federated Horizontal Pod Autoscalers (HPA)

FEATURE STATE: Kubernetes v1.16 alpha
This feature is currently in a alpha state, meaning:

  • The version names contain alpha (e.g. v1alpha1).
  • Might be buggy. Enabling the feature may expose bugs. Disabled by default.
  • Support for feature may be dropped at any time without notice.
  • The API may change in incompatible ways in a later software release without notice.
  • Recommended for use only in short-lived testing clusters, due to increased risk of bugs and lack of long-term support.

Deprecated

Use of Federation v1 is strongly discouraged. Federation V1 never achieved GA status and is no longer under active development. Documentation is for historical purposes only.

For more information, see the intended replacement, Kubernetes Federation v2.

This guide explains how to use federated horizontal pod autoscalers (HPAs) in the federation control plane.

HPAs in the federation control plane are similar to the traditional Kubernetes HPAs, and provide the same functionality. Creating an HPA targeting a federated object in the federation control plane ensures that the desired number of replicas of the target object are scaled across the registered clusters, instead of a single cluster. Also, the control plane keeps monitoring the status of each individual HPA in the federated clusters and ensures the workload replicas move where they are needed most by manipulating the min and max limits of the HPA objects in the federated clusters.

Before you begin

  • This guide assumes that you have a running Kubernetes Cluster Federation installation. If not, then head over to the federation admin guide to learn how to bring up a cluster federation (or have your cluster administrator do this for you). Other tutorials, such as Kelsey Hightower’s Federated Kubernetes Tutorial, might also help you create a Federated Kubernetes cluster.

  • You should also have a basic working knowledge of Kubernetes in general and HPAs in particular.

The federated HPA is an alpha feature. The API is not enabled by default on the federated API server. To use this feature, the user or the admin deploying the federation control plane needs to run the federated API server with option --runtime-config=api/all=true to enable all APIs, including alpha APIs. Additionally, the federated HPA only works when used with CPU utilization metrics.

Creating a federated HPA

The API for federated HPAs is 100% compatible with the API for traditional Kubernetes HPA. You can create an HPA by sending a request to the federation API server.

You can do that with kubectl by running:

cat <<EOF | kubectl --context=federation-cluster create -f -
apiVersion: autoscaling/v1
kind: HorizontalPodAutoscaler
metadata:
  name: php-apache
  namespace: default
spec:
  scaleTargetRef:
    apiVersion: apps/v1beta1
    kind: Deployment
    name: php-apache
  minReplicas: 1
  maxReplicas: 10
  targetCPUUtilizationPercentage: 50
EOF

The --context=federation-cluster flag tells kubectl to submit the request to the federation API server instead of sending it to a Kubernetes cluster.

Once a federated HPA is created, the federation control plane partitions and creates the HPA in all underlying Kubernetes clusters. As of Kubernetes V1.7, cluster selectors can also be used to restrict any federated object, including the HPAs in a subset of clusters.

You can verify the creation by checking each of the underlying clusters. For example, with a context named gce-asia-east1a configured in your client for your cluster in that zone:

kubectl --context=gce-asia-east1a get HPA php-apache

The HPA in the underlying clusters will match the federation HPA except in the number of min and max replicas. The federation control plane ensures that the sum of max replicas in each cluster matches the specified max replicas on the federated HPA object, and the sum of minimum replicas will be greater than or equal to the minimum specified on the federated HPA object.

Note: A particular cluster cannot have a minimum replica sum of 0.

Spreading HPA min and max replicas in underlying clusters

By default, first max replicas are spread equally in all the underlying clusters, then min replicas are distributed to those clusters that received their maximum value. This means that each cluster will get an HPA if the specified max replicas are greater than the total clusters participating in this federation, and some clusters will be skipped if specified max replicas are less than the total clusters participating in the federation.

For example: if you have 3 registered clusters and you create a federated HPA with spec.maxReplicas = 9, and spec.minReplicas = 2, then each HPA in the 3 clusters will get spec.maxReplicas=3 and spec.minReplicas = 1.

Currently the default distribution is only available on the federated HPA, but in the future, users preferences could also be specified to control and/or restrict this distribution.

Updating a federated HPA

You can update a federated HPA as you would update a Kubernetes HPA; however, for a federated HPA, you must send the request to the federation API server instead of sending it to a specific Kubernetes cluster. The Federation control plane ensures that whenever the federated HPA is updated, it updates the corresponding HPA in all underlying clusters to match it.

If your update includes a change in the number of replicas, the federation control plane will change the number of replicas in underlying clusters to ensure that the sum of the max and min replicas remains matched as specified in the previous section.

Deleting a federated HPA

You can delete a federated HPA as you would delete a Kubernetes HPA; however, for a federated HPA, you must send the request to the federation API server instead of to a specific Kubernetes cluster.

Note: For the federated resource to be deleted from all underlying clusters, cascading deletion should be used.

For example, you can do that using kubectl by running:

kubectl --context=federation-cluster delete HPA php-apache

Alternative ways to use federated HPA

To a federation user interacting with federated control plane (or simply federation), the interaction is almost identical to interacting with a normal Kubernetes cluster (but with a limited set of APIs that are federated). As both Deployments and HorizontalPodAutoscalers are now federated, kubectl commands like kubectl run and kubectl autoscale work on federation. Given this fact, the mechanism specified in horizontal pod autoscaler walkthrough will also work when used with federation. Care however will need to be taken that when generating load on a target deployment, it should be done against a specific federated cluster (or multiple clusters) not the federation.

Conclusion

The use of federated HPA is to ensure workload replicas move to the cluster(s) where they are needed most, or in other words where the load is beyond expected threshold. The federated HPA feature achieves this by manipulating the min and max replicas on the HPAs it creates in the federated clusters. It does not directly monitor the target object metrics from the federated clusters. It actually relies on the in-cluster HPA controllers to monitor the metrics and update relevant fields. The in-cluster HPA controller monitors the target pod metrics and updates the fields like desired replicas (after metrics based calculations) and current replicas (observing the current status of in cluster pods). The federated HPA controller, on the other hand, monitors only the cluster-specific HPA object fields and updates the min replica and max replica fields of those in cluster HPA objects, which have replicas matching thresholds.

For example, if a cluster has both desired replicas and current replicas the same as the max replicas, and averaged current CPU utilization still higher than the target CPU utilization (all of which are fields on local HPA object), then the target app in this cluster needs more replicas, and the scaling is currently restricted by max replicas set on this local HPA object. In such a scenario, the federated HPA controller scans all clusters and tries to find clusters which do not have such a condition (meaning the desired replicas are less than the max, and current averaged CPU utilization is lower then the threshold). If it finds such a cluster, it reduces the max replica on the HPA in this cluster and increases the max replicas on the HPA in the cluster which needed the replicas.

There are many other similar conditions which the federated HPA controller checks and moves the max replicas and min replicas around the local HPAs in federated clusters to eventually ensure that the replicas move (or remain) in the cluster(s) which need them.

For more information, see “federated HPA design proposal”.

Feedback