Autoscaling is the ability of your Gardener shoot cluster to add and remove nodes as needed, based on the workload you deploy to the Kubernetes cluster.
Autoscaling in Cleura Cloud is always horizontal rather than vertical, meaning Gardener adds or removes nodes of the pre-defined machine type (flavor), rather than changing the flavor of existing nodes in-place.
By default, Gardener launches 3 initial worker nodes, and limits your cluster to a maximum of 5 nodes.
You can modify these settings (Autoscaler Min and Autoscaler Max) when you create a new shoot cluster, or at any time thereafter.
How autoscaling works
Autoscaling is designed to be a “hands-off” cluster feature.
Once your cluster decides its resources are insufficient to manage the current workload, it adds nodes to handle it. This is called scale-out.
When the workload becomes less, it removes the extra nodes again. This is called scale-in.
Scale-out is triggered if any Pod on the cluster fails to be scheduled to an existing node.
This may be because the node is already overloaded (meaning its
PIDPressure node condition is active), or because a newly launched Pod is configured with a request that none of the existing nodes can meet.
Thus, scaling out is nearly immediate: as soon as a Pod scheduling failure occurs for one of these reasons, the cluster launches a new worker node — unless it is already running with the maximum number of nodes, as defined by Autoscaler Max.
Scale-in, in contrast, happens when a node’s utilization drops below 50% for a period of 30 minutes. Therefore, scaling in occurs in a delayed fashion. This is by design: if the dip in resource utilization is only temporary, sufficient worker node capacity is already available when it rebounds.
Autoscaling may sometimes not occur when you expect it to, including the following situations:
If a Pod spec contains a request that is impossible to meet even with scale-out, no autoscaling occurs. For example, if a Pod were to request 1TiB of memory, and the configured worker node flavor has less than that, then scale-out would not help: to run such a Pod, you would instead have to add a new worker group (with a larger flavor) to the cluster.
If a Deployment contains Pods with anti-affinity rules that restrict multiple replicas from running on the same node, and its number of running replicas is already equal to the current number of worker nodes in the group, then no scale-in is possible. For example, suppose a Deployment is running with 4 replicas of an anti-affinity Pod in a cluster with default Autoscaler Min and Autoscaler Max values. When at some point that cluster has 4 worker nodes, the fourth node will remain even if its utilization is permanently below 50%.