Karpenter is a metrics-driven autoscaler built for Kubernetes and can run in any Kubernetes cluster anywhere. It's performant, extensible, and can autoscale anything that implements the Kubernetes scale subresource.
This is an early stage, experimental project built with ❤️ and is available as a developer preview. We're excited you are here - jump in, let us know what you think. We welcome contributions.
We will learn about Karpenter's APIs, look at some sample configurations, and install Karpenter's Controller. Alternatively, you can dive right into the demo.
Karpenter defines three custom resources to configure autoscaling behavior.
HorizontalAutoscalers define your autoscaling policy. It's modeled closely after the HorizontalPodAutoscaler, but has been generalized to support autoscaling for arbitrary resources. HorizontalAutoscalers periodically query metrics configured by spec.metrics, compute an autoscaling decision controlled by spec.behavior, and adjust the replicas of their spec.scaleTargetRef. Unlike the HPA, Karpenter's HorizontalAutoscalers integrate directly with Prometheus and can use any promql response of type "instant vector" in their calculations.
MetricsProducers generate Prometheus metrics for commonly used autoscaling use cases. They periodically calculate a metric based on their configuration and expose it at a metrics endpoint that can be scraped by Prometheus. If you already have metrics you wish to use for autoscaling available in Prometheus, it is not necessary to define a Metrics Producer.
ScalableNodeGroups provide a minimal way to point a HorizontalAutoscaler's scaleTargetRef to a Cloud Provider's Node Group API. Kubernetes core does not define an abstraction for Node Group. Instead, Cloud Providers typically expose non-Kubernetes Node Group APIs. ScalableNodeGroups are a shim in front of these APIs that are limited to spec.replicas and status.replicas. It is not a replacement or wrapper for these APIs. If you're using a solution that provides a Kubernetes API (e.g. Kops or Cluster API), you can point the HorizontalAutoscaler's scaleTargetRef to these resources instead of a ScalableNodeGroup.
Learn more about the different ways to configure Karpenter's resources.
Follow the setup recommendations of your cloud provider.
Then install the controller.
CLOUD_PROVIDER=aws
VERSION=v0.1.0
kubectl apply -f https://raw.githubusercontent.com/awslabs/karpenter/main/releases/${CLOUD_PROVIDER}/${VERSION}.yaml
Karpenter is an early stage, experimental project that is currently maintained by AWS and available as a preview. We request that you do not use Karpenter for production workloads at this time. See details in our terms.
This project is licensed under the Apache-2.0 License.