> ## Documentation Index
> Fetch the complete documentation index at: https://docs.costgraph.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Supported node autoscalers

> Check which node autoscalers CostGraph recognizes, what it recommends for each, and what AutoPilot can change.

CostGraph recognizes the autoscaler behind each node pool. The autoscaler sets
the form of a node pool recommendation, whether AutoPilot can create pools, and
which options the [autoscaler choice insight](https://docs.costgraph.ai/costgraph/insights.md#autoscaler-choice)
compares. For the labels behind each pool type, refer to
[Supported node pools](https://docs.costgraph.ai/costgraph/operator/supported-node-pools.md).

AutoPilot sizes workload requests on every cluster, whatever its autoscaler.
It does not edit a pool that you created. Refer to
[Karpenter node pools](https://docs.costgraph.ai/costgraph/autopilot/karpenter.md#your-own-pools) for how
changes to your own pools reach you.

## Pool definitions

CostGraph reads node labels and annotations on every cluster. Karpenter, EKS
Auto Mode, AKS node auto-provisioning, and GKE compute classes also need their
pool definitions. These are NodePool or ComputeClass objects and the
`EC2NodeClass`, `AKSNodeClass`, or `OCINodeClass` objects they reference. These let CostGraph check instance
requirements, disruption settings, and node pod limits.

The operator reads these objects when `rbac.grantCustomResources` is `true`,
which is the default. Install or upgrade the chart with `helm install` or
`helm upgrade` so Helm discovers the cluster's custom resource API groups.
Until CostGraph has a pool definition, the node pool savings insight shows that
it has not been collected from the cluster yet.

## Supported autoscalers

Find your autoscaler to see how CostGraph recognizes it and what it recommends.

### In-cluster autoscalers

These autoscalers run as controllers in your cluster.

> [accordion] Karpenter
>
> * CostGraph recognizes each NodePool from the `karpenter.sh/nodepool` node label.
> * The node pool savings insight shows a NodePool manifest and a diff against your live NodePool. It flags consolidation that removes only empty nodes, a disruption budget that allows no node removals, and on-demand nodes in a pool that allows spot.
> * On Amazon EKS with self-managed Karpenter and the `karpenter.sh/v1` API, AutoPilot can create its own NodePools when splitting a pool saves money. This is off by default. Refer to [Karpenter node pools](https://docs.costgraph.ai/costgraph/autopilot/karpenter.md) to turn it on. AutoPilot does not create pools for Karpenter on other clouds.
> * The autoscaler choice insight compares Karpenter on Amazon EKS. A move to Karpenter requires spot interruption handling, from AWS Node Termination Handler or a Karpenter controller with an interruption queue.
> * The cluster needs the objects listed in [Pool definitions](#pool-definitions).

> [accordion] Cluster API
>
> * CostGraph recognizes each MachineSet from the annotations that Cluster API sets on its nodes.
> * The node pool savings insight lists the proposed instance types. Apply them through your MachineDeployment or infrastructure code.
> * AutoPilot does not change Cluster API resources. The autoscaler choice insight does not cover Cluster API clusters.
> * The cluster needs nothing beyond the operator.

> [accordion] Gardener
>
> * CostGraph recognizes worker pools from the `worker.gardener.cloud/pool` node label. On STACKIT nodes, the label identifies a STACKIT SKE worker pool.
> * The node pool savings insight lists the proposed instance types. Apply them through your Shoot specification or infrastructure code. Each STACKIT worker pool holds one machine type, so the insight suggests one pool per proposed type.
> * AutoPilot does not change Gardener worker pools. The autoscaler choice insight does not cover Gardener clusters.
> * The cluster needs nothing beyond the operator.

### Cloud-managed autoscalers

The cloud provider runs these autoscalers. CostGraph treats EKS managed node
groups, GKE Standard node pools, and AKS node pools as groups that the cluster
autoscaler scales. It does not check for the cluster autoscaler deployment.
AutoPilot does not create pools for any cloud-managed autoscaler.

> [accordion] Amazon EKS
>
> * Managed node groups with Cluster Autoscaler use the `eks.amazonaws.com/nodegroup` label. The node pool savings insight shows an `eksctl` cluster configuration, with on-demand and spot capacity as separate node groups.
> * EKS Auto Mode runs a managed Karpenter. CostGraph treats a Karpenter pool as EKS Auto Mode when its NodePool references a node class in the `eks.amazonaws.com` group. It also does so when the pool's nodes carry the `eks.amazonaws.com/compute-type: auto` label. The insight shows a NodePool manifest and a diff against your live NodePool.
> * EKS Auto Mode adds a management fee that CostGraph does not price yet. The autoscaler choice insight lists it without comparing it, and does not run on a cluster where any pool uses EKS Auto Mode.
> * EKS Auto Mode needs the objects listed in [Pool definitions](#pool-definitions).

> [accordion] Google Kubernetes Engine
>
> * Standard node pools with cluster autoscaler use the `cloud.google.com/gke-nodepool` label. Node auto-provisioning creates node pools whose names start with `nap-`. For both, the node pool savings insight lists the proposed instance types to create in your console or infrastructure code.
> * Custom compute classes use the `cloud.google.com/compute-class` label. The insight shows a ComputeClass manifest with the proposed machine priorities and a diff against your live class. It flags a class whose priorities are all spot and that does not fall back to on-demand, because pods wait when spot capacity runs short. Cost reports group compute class nodes under their underlying node pool.
> * GKE Autopilot bills pods by their requests. CostGraph identifies Autopilot nodes from the machine family that GKE reports for them. Autopilot adds a management fee that CostGraph does not price yet. The autoscaler choice insight lists it without comparing it, and does not run on a cluster where any pool runs on Autopilot.
> * A move to node auto-provisioning requires a cluster with no Windows nodes.
> * Compute classes need the objects listed in [Pool definitions](#pool-definitions).

> [accordion] Microsoft Azure AKS
>
> * Node pools with cluster autoscaler use the `kubernetes.azure.com/agentpool` label. The node pool savings insight lists the proposed instance types. An AKS node pool keeps one instance type, so the insight suggests creating the proposed pool and draining the current one.
> * Node auto-provisioning runs Karpenter, so CostGraph recognizes its pools from the `karpenter.sh/nodepool` label on AKS nodes. The insight shows a NodePool manifest and a diff against your live NodePool.
> * A move to node auto-provisioning requires a cluster with no Windows nodes.
> * Node auto-provisioning needs the objects listed in [Pool definitions](#pool-definitions).

## Next steps

Choose a guide to compare autoscalers or set up pools that AutoPilot manages.

> [card] [Autoscaler choice insight](https://docs.costgraph.ai/costgraph/insights.md#autoscaler-choice)
>
> Compare what each autoscaler would cost on your workloads.

> [card] [Karpenter node pools](https://docs.costgraph.ai/costgraph/autopilot/karpenter.md)
>
> Let AutoPilot create and manage its own NodePools.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.