Autoscaling

Autoscaling is a way of changing the size of a node group, number of pods, or the amount of resources allocated to each pod based on resource requests for pods running on this group's nodes. Autoscaling is available in Kubernetes starting from version 1.15.

In a Managed Service for Kubernetes cluster, autoscaling can be implemented in three ways:

• Cluster autoscaling (Cluster Autoscaler). Managed Service for Kubernetes monitors workloads on the nodes and updates the number of nodes within specified limits as required.
• Horizontal pod scaling (Horizontal Pod Autoscaler). Kubernetes dynamically changes the number of pods running on each node in the group.
• Vertical pod scaling (Vertical Pod Autoscaler). When workloads increase, Kubernetes allocates additional resources to each pod within the set limits.

You can employ several types of autoscaling in the same cluster. However, using Horizontal Pod Autoscaler and Vertical Pod Autoscaler together is not recommended.

Cluster autoscalingCluster autoscaling

Cluster Autoscaler automatically modifies the number of nodes in a group depending on your workloads.

When creating a node group, select an autoscaling type and set the minimum, maximum, and initial number of nodes in the group. Kubernetes will periodically check the pod status and workloads on the nodes, adjusting the group size as required:

• If pods cannot be assigned due to a lack of vCPUs or RAM on the existing nodes, the number of nodes in the group will gradually increase to the specified maximum.
• If a workload on nodes is low, and all pods can be assigned to fewer nodes in a group, the number of nodes in the group will gradually decrease to the specified minimum. If pods on the node cannot be evicted within the specified period of time (7 minutes), the node is forced to stop. The timeout cannot be changed.

Cluster Autoscaler activation is only available when creating a node group. Cluster Autoscaler is managed on the Managed Service for Kubernetes side.

For more information, see these Kubernetes guides:

• Cluster Autoscaler description
• Default parameters

See also Questions and answers about node group autoscaling in Managed Service for Kubernetes.

Horizontal pod autoscalingHorizontal pod autoscaling

When using horizontal pod scaling, Kubernetes changes the number of pods depending on vCPU workload.

When creating a Horizontal Pod Autoscaler, specify the following parameters:

• vCPU load average percentage for each pod.
• Minimum and maximum number of pod replicas.

Horizontal pod autoscaling is available for the following controllers:

• Deployment.
• StatefulSet.
• ReplicaSet.

For more on Horizontal Pod Autoscaler, see this Kubernetes guide.

Vertical pod autoscalingVertical pod autoscaling

Kubernetes uses the limits parameters to restrict resources allocated for each application. A pod exceeding the vCPU limit will trigger CPU throttling. A pod exceeding the RAM limit will be stopped.

If required, Vertical Pod Autoscaler allocates additional vCPU and RAM resources to pods.

When creating a Vertical Pod Autoscaler, set the autoscaling mode in the specification:

• updateMode: "Off" for Vertical Pod Autoscaler to provide recommendations on managing pod resources without modifying them.
• updateMode: "Initial" for Vertical Pod Autoscaler to apply recommendations only when creating pods.
• updateMode: "Recreate" for Vertical Pod Autoscaler to recreate pods with updated resource values in case of a serious discrepancy between the current requests and recommendations.
• updateMode: "InPlaceOrRecreate" for Vertical Pod Autoscaler to attempt updating requests and resource limits first, without restarting the pod. If such an update is not possible, the pod will be recreated as in the Recreate mode. For more information, see Resize CPU and Memory Resources assigned to Containers.

For more on Vertical Pod Autoscaler, see this Kubernetes guide.

Use casesUse cases

• Horizontal scaling of an application in a cluster
• Vertical scaling of an application in a cluster
• Deploying and load testing a scalable gRPC service