Federated Pod Autoscaler

Requirements & Design Document

irfan.rehman@huawei.com, quinton.hoole@huawei.com

Use cases

1 – Users can schedule replicas of same application, across the federated clusters, using replicaset (or deployment). Users however further might need to let the replicas be scaled independently in each cluster, depending on the current usage metrics of the replicas; including the CPU, memory and application defined custom metrics.

2 - As stated in the previous use case, a federation user schedules replicas of same application, into federated clusters and subsequently creates a horizontal pod autoscaler targeting the object responsible for the replicas. User would want the auto-scaling to continue based on the in-cluster metrics, even if for some reason, there is an outage at federation level. User (or other users) should still be able to access the deployed application into all federated clusters. Further, if the load on the deployed app varies, the autoscaler should continue taking care of scaling the replicas for a smooth user experience.

3 - A federation that consists of an on-premise cluster and a cluster running in a public cloud has a user workload (eg. deployment or rs) preferentially running in the on-premise cluster. However if there are spikes in the app usage, such that the capacity in the on-premise cluster is not sufficient, the workload should be able to get scaled beyond the on-premise cluster boundary and into the other clusters which are part of this federation.

Please refer to some additional use cases, which partly led to the derivation of the above use case, and are listed in the glossary section of this document.

User workflow

User wants to schedule a set of common workload across federated clusters. He creates a replicaset or a deployment to schedule the workload (with or without preferences). The federation then distributes the replicas of the given workload into the federated clusters. As the user at this point is unaware of the exact usage metrics of the individual pods created in the federated clusters, he creates an HPA into the federation, providing metric parameters to be used in the scale request for a resource. It is now the responsibility of this HPA to monitor the relevant resource metrics and the scaling of the pods per cluster then is controlled by the associated HPA.

Alternative approaches

Design Alternative 1

Make the autoscaling resource available and implement support for horizontalpodautoscalers objects at federation. The HPA API resource will need to be exposed at the federation level, which can follow the version similar to one implemented in the latest k8s cluster release.

Once the HPA object is created at federation, the federation controller creates and monitors a similar HPA object (partitioning the min and max values) in each of the federated clusters. Based on the metadata in spec of the HPA describing the scaleTargetRef, the HPA will be applied on the already existing target objects. If the target object is not present in the cluster (either because, its not created until now, or deleted for some reason), the HPA will still exist but no action will be taken. The HPA's action will become applicable when the target object is created in the given cluster anytime in future. Also as stated already the federation controller will need to partition the min and max values appropriately into the federated clusters among the HPA objects such that the total of min and that of max replicas satisfies the constraints specified by the user at federation. The point of control over the scaling of replicas will lie locally with the federated hpa controller. The federated controller will however watch the cluster local HPAs wrt current replicas of the target objects and will do intelligent dynamic adjustments of min and max values of the HPA replicas across the clusters based on the run time conditions.

The federation controller by default will distribute the min and max replicas of the HPA equally among all clusters. The min values will first be distributed such that any cluster into which the replicas are distributed does not get a min replicas lesser than 1. This means that HPA can actually be created in lesser number of ready clusters then available in federation. Once this distribution happens, the max replicas of the hpa will be distributed across all those clusters into which the HPA needs to be created. The default distribution can be overridden using the annotations on the HPA object, very similar to the annotations on federated replicaset object as described here.

One of the points to note here is that, doing this brings a two point control on number of replicas of the target object, one by the federated target object (rs or deployment) and other by the hpa local to the federated cluster. Solution to which is discussed in the following section. Another additional note here is that, the preferences would consider use of only minreplicas and maxreplicas in this phase of implementation and weights will be discarded for this alternative design.

Rebalancing of workload replicas and control over the same.

The current implementation of federated replicasets (and deployments) first distributes the replicas into underlying clusters and then monitors the status of the pods in each cluster. In case there are clusters which have active pods lesser than what federation reconciler desires, federation control plane will trigger creation of the missing pods (which federation considers missing), or in other case would trigger removal of pods, if the control plane considers that the given cluster has more pods than needed. This is something which counters the role of HPA in individual cluster. To handle this, the knowledge that HPA is active separately targeting this object has to be percolated to the federation control plane monitoring the individual replicas such that, the federation control plane stops reconciling the replicas in the individual clusters. In other words the link between the HPA wrt to the corresponding objects will need to be maintained and if an HPA is active, other federation controllers (aka replicaset and deployment controllers) reconcile process, would stop updating and/or rebalancing the replicas in and across the underlying clusters. The reconcile of the objects (rs or deployment) would still continue, to handle the scenario of the object missing from any given federated cluster. The mechanism to achieve this behaviour shall be as below: - User creates a workload object (for example rs) in federation. - User then creates an HPA object in federation (this step and the previous step can follow either order of creation). - The rs as an object will exist in federation control plane with or without the user preferences and/or cluster selection annotations. - The HPA controller will first evaluate which cluster(s) get the replicas and which don’t (if any). This list of clusters will be a subset of the cluster selector already applied on the hpa object. - The HPA controller will apply this list on the federated rs object as the cluster selection annotation overriding the user provided preferences (if any). The control over the placement of workload replicas and the add on preferences will thus lie completely with the HPA objects. This is an important assumption that the user of these federated objects interacting with each other should be aware of; and if the user needs to place replicas in specific clusters, together with workload autoscaling he/she should apply these preferences on the HPA object. Any preferences applied on the workload object (rs or deployment) will be overridden. - The target workload object (for example rs) replicas will be kept unchanged in the cluster which already has the replicas, will be created with one replica if the particular cluster does not have the same and HPA calculation resulted in some replicas for that cluster and deleted from the clusters which has the replicas and the federated HPA calculations result in no replicas for that particular cluster. - The desired replicas per cluster as per the federated HPA dynamic rebalance mechanism, elaborated in the next section, will be set on individual clusters local HPA, which in turn will set the same on the target local object.

Dynamic HPA min/max rebalance

The proposal in this section can be used to improve the distribution of replicas across the clusters such that there are more replicas in those clusters, where they are needed more. The federation hpa controller will monitor the status of the local HPAs in the federated clusters and update the min and/or max values set on the local HPAs as below (assuming that all previous steps are done and local HPAs in federated clusters are active):

  1. At some point, one or more of the cluster HPA's hit the upper limit of their allowed scaling such that DesiredReplicas == MaxReplicas; Or more appropriately CurrentReplicas == DesiredReplicas == MaxReplicas.

  2. If the above is observed the Federation HPA tries to transfer allocation of MaxReplicas from clusters where it is not needed (DesiredReplicas < MaxReplicas) or where it cannot be used, e.g. due to capacity constraints (CurrentReplicas < DesiredReplicas <= MaxReplicas) to the clusters which have reached their upper limit (1 above).

  3. It will be taken care that the MaxReplica does not become lesser than MinReplica in any of the clusters in this redistribution. Additionally if the usage of the same could be established, MinReplicas can also be distributed as in 4 below.

  4. An exactly similar approach can also be applied to MinReplicas of the local HPAs, so as to reduce the min from those clusters, where
    CurrentReplicas == DesiredReplicas == MinReplicas and the observed average resource metric usage (on the HPA) is lesser then a given threshold, to those clusters, where the DesiredReplicas > MinReplicas.

However, as stated in 3 above, the approach of distribution will first be implemented only for MaxReplicas to establish it utility, before implementing the same for MinReplicas.

Design Alternative 2

Same as the previous one, the API will need to be exposed at federation.

However, when the request to create HPA is sent to federation, federation controller will not create the HPA into the federated clusters. The HPA object will reside in the federation API server only. The federation controller will need to get a metrics client to each of the federated clusters and collect all the relevant metrics periodically from all those clusters. The federation controller will further calculate the current average metrics utilisation across all clusters (using the collected metrics) of the given target object and calculate the replicas globally to attain the target utilisation as specified in the federation HPA. After arriving at the target replicas, the target replica number is set directly on the target object (replicaset, deployment, ..) using its scales sub-resource at federation. It will be left to the actual target object controller (for example RS controller) to distribute the replicas accordingly into the federated clusters. The point of control over the scaling of replicas will lie completely with the federation controllers.

Algorithm (for alternative 2)

Federated HPA (FHPA), from every cluster gets:

  • avg_i average metric value (like CPU utilization) for all pods matching the deployment/rs selector.
  • count_i number of replicas that were used to calculate the average.

To calculate the target number of replicas HPA calculates the sum of all metrics from all clusters:

sum(avg_i * count_i) and divides it by target metric value. The target replica count (validated against HPA min/max and thresholds) is set on Federated Deployment/replica set. So the deployment has the correct number of replicas (that should match the desired metric value) and provides all of the rebalancing/failover mechanisms.

Further, this can be expanded such that FHPA places replicas where they are needed the most (in cluster that have the most traffic). For that FHPA would play with weights in Federated Deployment. Each cluster will get the weight of 100 * avg_i/sum(avg_i). Weights hint Federated Deployment where to put replicas. But they are only hints so if placing a replica in the desired cluster is not possible then it will be placed elsewhere, what is probably better than not having the replica at all.

Other Scenario

Other scenario, for example rolling updates (when user updates the deployment or RS), recreation of the object (when user specifies the strategy as recreate while updating the object), will continue to be handled the way they are handled in an individual k8s cluster. Additionally there is a shortcoming in the current implementation of the federated deployments rolling update. There is an existing proposal as part of the federated deployment design doc. Given it is implemented, the rolling updates for a federated deployment while a federated HPA is active on the same object will also work fine.


The design alternative 2 has the following major drawbacks, which are sufficient to discard it as a probable implementation option: - This option needs the federation control plane controller to collect metrics data from each cluster, which is an overhead with increasing gravity of the problem with increasing number of federated clusters, in a given federation. - The monitoring and update of objects which are targeted by the federated HPA object (when needed) for a particular federated cluster would stop if for whatever reasons the network link between the federated cluster and federation control plane is severed. A bigger problem can happen in case of an outage of the federation control plane altogether.

In Design Alternative 1 the autoscaling of replicas will continue, even if a given cluster gets disconnected from federation or in case of the federation control plane outage. This would happen because the local HPAs with the last know maxreplica and minreplicas would exist in the local clusters. Additionally in this alternative there is no need of collection and processing of the pod metrics for the target object from each individual cluster. This document proposes to use design alternative 1 as the preferred implementation.


These use cases are specified using the terminology partly specific to telecom products/platforms:

1 - A telecom service provider has a large number of base stations, for a particular region, each with some set of virtualized resources each running some specific network functions. In a specific scenario the resources need to be treated logically separate (thus making large number of smaller clusters), but still a very similar workload needs to be deployed on each cluster (network function stacks, for example).

2 - In one of the architectures, the IOT matrix has IOT gateways, which aggregate a large number of IOT sensors in a small area (for example a shopping mall). The IOT gateway is envisioned as a virtualized resource, and in some cases multiple such resources need aggregation, each forming a small cluster. Each of these clusters might run very similar functions, but will independently scale based on the demand of that area.

3 - A telecom service provider has a large number of base stations, each with some set of virtualized resources, and each running specific network functions and each specifically catering to different network abilities (2g, 3g, 4g, etc). Each of these virtualized base stations, make small clusters and can cater to specific network abilities, such that one can cater to one or more network abilities. At a given point of time there would be some number of end user agents (cell phones) associated with each, and these UEs can come and go within the range of each. While the UEs move, a more centralized entity (read federation) needs to make a decision as to which exact base station cluster is suitable and with needed resources to handle the incoming UEs.