Skip to main content

Component: eks/karpenter

This component provisions Karpenter on an EKS cluster. It requires at least version 0.32.0 of Karpenter, though you are encouraged to use the latest version.

Usage

Stack Level: Regional

These instructions assume you are provisioning 2 EKS clusters in the same account and region, named "blue" and "green", and alternating between them. If you are only using a single cluster, you can ignore the "blue" and "green" references and remove the metadata block from the karpenter module.

components:
terraform:
# Base component of all `karpenter` components
eks/karpenter:
metadata:
type: abstract
vars:
enabled: true
eks_component_name: "eks/cluster"
name: "karpenter"
# https://github.com/aws/karpenter/tree/main/charts/karpenter
chart_repository: "oci://public.ecr.aws/karpenter"
chart: "karpenter"
chart_version: "v0.36.0"
# Enable Karpenter to get advance notice of spot instances being terminated
# See https://karpenter.sh/docs/concepts/#interruption
interruption_handler_enabled: true
resources:
limits:
cpu: "300m"
memory: "1Gi"
requests:
cpu: "100m"
memory: "512Mi"
cleanup_on_fail: true
atomic: true
wait: true
rbac_enabled: true
# "karpenter-crd" can be installed as an independent helm chart to manage the lifecycle of Karpenter CRDs
crd_chart_enabled: true
crd_chart: "karpenter-crd"
# replicas set the number of Karpenter controller replicas to run
replicas: 2
# "settings" controls a subset of the settings for the Karpenter controller regarding batch idle and max duration.
# you can read more about these settings here: https://karpenter.sh/docs/reference/settings/
settings:
batch_idle_duration: "1s"
batch_max_duration: "10s"
# The logging settings for the Karpenter controller
logging:
enabled: true
level:
controller: "info"
global: "info"
webhook: "error"

Provision Karpenter on EKS cluster

Here we describe how to provision Karpenter on an EKS cluster. We will be using the plat-ue2-dev stack as an example.

Provision Service-Linked Roles for EC2 Spot and EC2 Spot Fleet

Note: If you want to use EC2 Spot for the instances launched by Karpenter, you may need to provision the following Service-Linked Role for EC2 Spot:

  • Service-Linked Role for EC2 Spot

This is only necessary if this is the first time you're using EC2 Spot in the account. Since this is a one-time operation, we recommend you do this manually via the AWS CLI:

aws --profile <namespace>-<tenamt>-gbl-<stage>-admin iam create-service-linked-role --aws-service-name spot.amazonaws.com

Note that if the Service-Linked Roles already exist in the AWS account (if you used EC2 Spot or Spot Fleet before), and you try to provision them again, you will see the following errors:

An error occurred (InvalidInput) when calling the CreateServiceLinkedRole operation:
Service role name AWSServiceRoleForEC2Spot has been taken in this account, please try a different suffix

For more details, see:

The process of provisioning Karpenter on an EKS cluster consists of 3 steps.

1. Provision EKS IAM Role for Nodes Launched by Karpenter

VPC assumptions being made

We assume you've already created a VPC using our VPC component and have private subnets already set up. The Karpenter node pools will be launched in the private subnets.

EKS IAM Role for Nodes launched by Karpenter are provisioned by the eks/cluster component. (EKS can also provision a Fargate Profile for Karpenter, but deploying Karpenter to Fargate is not recommended.):

components:
terraform:
eks/cluster-blue:
metadata:
component: eks/cluster
inherits:
- eks/cluster
vars:
karpenter_iam_role_enabled: true
Authorization
  • The AWS Auth API for EKS is used to authorize the Karpenter controller to interact with the EKS cluster.

Karpenter is installed using a Helm chart. The Helm chart installs the Karpenter controller and a webhook pod as a Deployment that needs to run before the controller can be used for scaling your cluster. We recommend a minimum of one small node group with at least one worker node.

As an alternative, you can run these pods on EKS Fargate by creating a Fargate profile for the karpenter namespace. Doing so will cause all pods deployed into this namespace to run on EKS Fargate. Do not run Karpenter on a node that is managed by Karpenter.

See Run Karpenter Controller... for more details.

We provision IAM Role for Nodes launched by Karpenter because they must run with an Instance Profile that grants permissions necessary to run containers and configure networking.

We define the IAM role for the Instance Profile in components/terraform/eks/cluster/controller-policy.tf.

Note that we provision the EC2 Instance Profile for the Karpenter IAM role in the components/terraform/eks/karpenter component (see the next step).

Run the following commands to provision the EKS Instance Profile for Karpenter and the IAM role for instances launched by Karpenter on the blue EKS cluster and add the role ARNs to the EKS Auth API:

atmos terraform plan eks/cluster-blue -s plat-ue2-dev
atmos terraform apply eks/cluster-blue -s plat-ue2-dev

For more details, refer to:

2. Provision karpenter component

In this step, we provision the components/terraform/eks/karpenter component, which deploys the following resources:

  • Karpenter CustomerResourceDefinitions (CRDs) using the Karpenter CRD Chart and the helm_release Terraform resource
  • Karpenter Kubernetes controller using the Karpenter Helm Chart and the helm_release Terraform resource
  • EKS IAM role for Kubernetes Service Account for the Karpenter controller (with all the required permissions)
  • An SQS Queue and Event Bridge rules for handling Node Interruption events (i.e. Spot)

Create a stack config for the blue Karpenter component in stacks/catalog/eks/clusters/blue.yaml:

eks/karpenter-blue:
metadata:
component: eks/karpenter
inherits:
- eks/karpenter
vars:
eks_component_name: eks/cluster-blue

Run the following commands to provision the Karpenter component on the blue EKS cluster:

atmos terraform plan eks/karpenter-blue -s plat-ue2-dev
atmos terraform apply eks/karpenter-blue -s plat-ue2-dev

3. Provision karpenter-node-pool component

In this step, we provision the components/terraform/eks/karpenter-node-pool component, which deploys Karpenter NodePools using the kubernetes_manifest resource.

Why use a separate component for NodePools?

We create the NodePools as a separate component since the CRDs for the NodePools are created by the Karpenter component. This helps manage dependencies.

First, create an abstract component for the eks/karpenter-node-pool component:

components:
terraform:
eks/karpenter-node-pool:
metadata:
type: abstract
vars:
enabled: true
# Disabling Manifest Experiment disables stored metadata with Terraform state
# Otherwise, the state will show changes on all plans
helm_manifest_experiment_enabled: false
node_pools:
default:
# Whether to place EC2 instances launched by Karpenter into VPC private subnets. Set it to `false` to use public subnets
private_subnets_enabled: true
# You can use disruption to set the maximum instance lifetime for the EC2 instances launched by Karpenter.
# You can also configure how fast or slow Karpenter should add/remove nodes.
# See more: https://karpenter.sh/v0.36/concepts/disruption/
disruption:
max_instance_lifetime: "336h" # 14 days
# Taints can be used to prevent pods without the right tolerations from running on this node pool.
# See more: https://karpenter.sh/v0.36/concepts/nodepools/#taints
taints: []
total_cpu_limit: "1k"
# Karpenter node pool total memory limit for all pods running on the EC2 instances launched by Karpenter
total_memory_limit: "1200Gi"
# Set acceptable (In) and unacceptable (Out) Kubernetes and Karpenter values for node provisioning based on
# Well-Known Labels and cloud-specific settings. These can include instance types, zones, computer architecture,
# and capacity type (such as AWS spot or on-demand).
# See https://karpenter.sh/v0.36/concepts/nodepools/#spectemplatespecrequirements for more details
requirements:
- key: "karpenter.sh/capacity-type"
operator: "In"
# See https://karpenter.sh/docs/concepts/nodepools/#capacity-type
# Allow fallback to on-demand instances when spot instances are unavailable
# By default, Karpenter uses the "price-capacity-optimized" allocation strategy
# https://aws.amazon.com/blogs/compute/introducing-price-capacity-optimized-allocation-strategy-for-ec2-spot-instances/
# It is currently not configurable, but that may change in the future.
# See https://github.com/aws/karpenter-provider-aws/issues/1240
values:
- "on-demand"
- "spot"
- key: "kubernetes.io/os"
operator: "In"
values:
- "linux"
- key: "kubernetes.io/arch"
operator: "In"
values:
- "amd64"
# The following two requirements pick instances such as c3 or m5
- key: karpenter.k8s.aws/instance-category
operator: In
values: ["c", "m", "r"]
- key: karpenter.k8s.aws/instance-generation
operator: Gt
values: ["2"]

Now, create the stack config for the blue Karpenter NodePool component in stacks/catalog/eks/clusters/blue.yaml:

eks/karpenter-node-pool/blue:
metadata:
component: eks/karpenter-node-pool
inherits:
- eks/karpenter-node-pool
vars:
eks_component_name: eks/cluster-blue

Finally, run the following commands to deploy the Karpenter NodePools on the blue EKS cluster:

atmos terraform plan eks/karpenter-node-pool/blue -s plat-ue2-dev
atmos terraform apply eks/karpenter-node-pool/blue -s plat-ue2-dev

Node Interruption

Karpenter also supports listening for and responding to Node Interruption events. If interruption handling is enabled, Karpenter will watch for upcoming involuntary interruption events that would cause disruption to your workloads. These interruption events include:

  • Spot Interruption Warnings
  • Scheduled Change Health Events (Maintenance Events)
  • Instance Terminating Events
  • Instance Stopping Events
Interruption Handler vs. Termination Handler

The Node Interruption Handler is not the same as the Node Termination Handler. The latter is always enabled and cleanly shuts down the node in 2 minutes in response to a Node Termination event. The former gets advance notice that a node will soon be terminated, so it can have 5-10 minutes to shut down a node.

For more details, see refer to the Karpenter docs and FAQ

To enable Node Interruption handling, set var.interruption_handler_enabled to true. This will create an SQS queue and a set of Event Bridge rules to deliver interruption events to Karpenter.

Custom Resource Definition (CRD) Management

Karpenter ships with a few Custom Resource Definitions (CRDs). In earlier versions of this component, when installing a new version of the karpenter helm chart, CRDs were not be upgraded at the same time, requiring manual steps to upgrade CRDs after deploying the latest chart. However Karpenter now supports an additional, independent helm chart for CRD management. This helm chart, karpenter-crd, can be installed alongside the karpenter helm chart to automatically manage the lifecycle of these CRDs.

To deploy the karpenter-crd helm chart, set var.crd_chart_enabled to true. (Installing the karpenter-crd chart is recommended. var.crd_chart_enabled defaults to false to preserve backward compatibility with older versions of this component.)

Troubleshooting

For Karpenter issues, checkout the Karpenter Troubleshooting Guide

References

For more details on the CRDs, see:

Requirements

NameVersion
terraform>= 1.3.0
aws>= 4.9.0
helm>= 2.0
kubernetes>= 2.7.1, != 2.21.0

Providers

NameVersion
aws>= 4.9.0

Modules

NameSourceVersion
ekscloudposse/stack-config/yaml//modules/remote-state1.5.0
iam_roles../../account-map/modules/iam-rolesn/a
karpentercloudposse/helm-release/aws0.10.1
karpenter_crdcloudposse/helm-release/aws0.10.1
thiscloudposse/label/null0.25.0

Resources

NameType
aws_cloudwatch_event_rule.interruption_handlerresource
aws_cloudwatch_event_target.interruption_handlerresource
aws_sqs_queue.interruption_handlerresource
aws_sqs_queue_policy.interruption_handlerresource
aws_eks_cluster_auth.eksdata source
aws_iam_policy_document.interruption_handlerdata source
aws_partition.currentdata source

Inputs

NameDescriptionTypeDefaultRequired
additional_tag_mapAdditional key-value pairs to add to each map in tags_as_list_of_maps. Not added to tags or id.
This is for some rare cases where resources want additional configuration of tags
and therefore take a list of maps with tag key, value, and additional configuration.
map(string){}no
atomicIf set, installation process purges chart on fail. The wait flag will be set automatically if atomic is usedbooltrueno
attributesID element. Additional attributes (e.g. workers or cluster) to add to id,
in the order they appear in the list. New attributes are appended to the
end of the list. The elements of the list are joined by the delimiter
and treated as a single ID element.
list(string)[]no
chartChart name to be installed. The chart name can be local path, a URL to a chart, or the name of the chart if repository is specified. It is also possible to use the <repository>/<chart> format here if you are running Terraform on a system that the repository has been added to with helm repo add but this is not recommendedstringn/ayes
chart_descriptionSet release description attribute (visible in the history)stringnullno
chart_repositoryRepository URL where to locate the requested chartstringn/ayes
chart_valuesAdditional values to yamlencode as helm_release valuesany{}no
chart_versionSpecify the exact chart version to install. If this is not specified, the latest version is installedstringnullno
cleanup_on_failAllow deletion of new resources created in this upgrade when upgrade failsbooltrueno
contextSingle object for setting entire context at once.
See description of individual variables for details.
Leave string and numeric variables as null to use default value.
Individual variable settings (non-null) override settings in context object,
except for attributes, tags, and additional_tag_map, which are merged.
any
{
"additional_tag_map": {},
"attributes": [],
"delimiter": null,
"descriptor_formats": {},
"enabled": true,
"environment": null,
"id_length_limit": null,
"label_key_case": null,
"label_order": [],
"label_value_case": null,
"labels_as_tags": [
"unset"
],
"name": null,
"namespace": null,
"regex_replace_chars": null,
"stage": null,
"tags": {},
"tenant": null
}
no
crd_chartThe name of the Karpenter CRD chart to be installed, if var.crd_chart_enabled is set to true.string"karpenter-crd"no
crd_chart_enabledkarpenter-crd can be installed as an independent helm chart to manage the lifecycle of Karpenter CRDs. Set to true to install this CRD helm chart before the primary karpenter chart.boolfalseno
delimiterDelimiter to be used between ID elements.
Defaults to - (hyphen). Set to "" to use no delimiter at all.
stringnullno
descriptor_formatsDescribe additional descriptors to be output in the descriptors output map.
Map of maps. Keys are names of descriptors. Values are maps of the form
{<br/> format = string<br/> labels = list(string)<br/>}
(Type is any so the map values can later be enhanced to provide additional options.)
format is a Terraform format string to be passed to the format() function.
labels is a list of labels, in order, to pass to format() function.
Label values will be normalized before being passed to format() so they will be
identical to how they appear in id.
Default is {} (descriptors output will be empty).
any{}no
eks_component_nameThe name of the eks componentstring"eks/cluster"no
enabledSet to false to prevent the module from creating any resourcesboolnullno
environmentID element. Usually used for region e.g. 'uw2', 'us-west-2', OR role 'prod', 'staging', 'dev', 'UAT'stringnullno
helm_manifest_experiment_enabledEnable storing of the rendered manifest for helm_release so the full diff of what is changing can been seen in the planboolfalseno
id_length_limitLimit id to this many characters (minimum 6).
Set to 0 for unlimited length.
Set to null for keep the existing setting, which defaults to 0.
Does not affect id_full.
numbernullno
interruption_handler_enabledIf true, deploy a SQS queue and Event Bridge rules to enable interruption handling by Karpenter.
https://karpenter.sh/docs/concepts/disruption/#interruption
booltrueno
interruption_queue_message_retentionThe message retention in seconds for the interruption handler SQS queue.number300no
kube_data_auth_enabledIf true, use an aws_eks_cluster_auth data source to authenticate to the EKS cluster.
Disabled by kubeconfig_file_enabled or kube_exec_auth_enabled.
boolfalseno
kube_exec_auth_aws_profileThe AWS config profile for aws eks get-token to usestring""no
kube_exec_auth_aws_profile_enabledIf true, pass kube_exec_auth_aws_profile as the profile to aws eks get-tokenboolfalseno
kube_exec_auth_enabledIf true, use the Kubernetes provider exec feature to execute aws eks get-token to authenticate to the EKS cluster.
Disabled by kubeconfig_file_enabled, overrides kube_data_auth_enabled.
booltrueno
kube_exec_auth_role_arnThe role ARN for aws eks get-token to usestring""no
kube_exec_auth_role_arn_enabledIf true, pass kube_exec_auth_role_arn as the role ARN to aws eks get-tokenbooltrueno
kubeconfig_contextContext to choose from the Kubernetes config file.
If supplied, kubeconfig_context_format will be ignored.
string""no
kubeconfig_context_formatA format string to use for creating the kubectl context name when
kubeconfig_file_enabled is true and kubeconfig_context is not supplied.
Must include a single %s which will be replaced with the cluster name.
string""no
kubeconfig_exec_auth_api_versionThe Kubernetes API version of the credentials returned by the exec auth pluginstring"client.authentication.k8s.io/v1beta1"no
kubeconfig_fileThe Kubernetes provider config_path setting to use when kubeconfig_file_enabled is truestring""no
kubeconfig_file_enabledIf true, configure the Kubernetes provider with kubeconfig_file and use that kubeconfig file for authenticating to the EKS clusterboolfalseno
label_key_caseControls the letter case of the tags keys (label names) for tags generated by this module.
Does not affect keys of tags passed in via the tags input.
Possible values: lower, title, upper.
Default value: title.
stringnullno
label_orderThe order in which the labels (ID elements) appear in the id.
Defaults to ["namespace", "environment", "stage", "name", "attributes"].
You can omit any of the 6 labels ("tenant" is the 6th), but at least one must be present.
list(string)nullno
label_value_caseControls the letter case of ID elements (labels) as included in id,
set as tag values, and output by this module individually.
Does not affect values of tags passed in via the tags input.
Possible values: lower, title, upper and none (no transformation).
Set this to title and set delimiter to "" to yield Pascal Case IDs.
Default value: lower.
stringnullno
labels_as_tagsSet of labels (ID elements) to include as tags in the tags output.
Default is to include all labels.
Tags with empty values will not be included in the tags output.
Set to [] to suppress all generated tags.
Notes:
The value of the name tag, if included, will be the id, not the name.
Unlike other null-label inputs, the initial setting of labels_as_tags cannot be
changed in later chained modules. Attempts to change it will be silently ignored.
set(string)
[
"default"
]
no
loggingA subset of the logging settings for the Karpenter controller
object({
enabled = optional(bool, true)
level = optional(object({
controller = optional(string, "info")
global = optional(string, "info")
webhook = optional(string, "error")
}), {})
})
{}no
nameID element. Usually the component or solution name, e.g. 'app' or 'jenkins'.
This is the only ID element not also included as a tag.
The "name" tag is set to the full id string. There is no tag with the value of the name input.
stringnullno
namespaceID element. Usually an abbreviation of your organization name, e.g. 'eg' or 'cp', to help ensure generated IDs are globally uniquestringnullno
rbac_enabledEnable/disable RBACbooltrueno
regex_replace_charsTerraform regular expression (regex) string.
Characters matching the regex will be removed from the ID elements.
If not set, "/[^a-zA-Z0-9-]/" is used to remove all characters other than hyphens, letters and digits.
stringnullno
regionAWS Regionstringn/ayes
replicasThe number of Karpenter controller replicas to runnumber2no
resourcesThe CPU and memory of the deployment's limits and requests
object({
limits = object({
cpu = string
memory = string
})
requests = object({
cpu = string
memory = string
})
})
n/ayes
settingsA subset of the settings for the Karpenter controller.
Some settings are implicitly set by this component, such as clusterName and
interruptionQueue. All settings can be overridden by providing a settings
section in the chart_values variable. The settings provided here are the ones
mostly likely to be set to other than default values, and are provided here for convenience.
object({
batch_idle_duration = optional(string, "1s")
batch_max_duration = optional(string, "10s")
})
{}no
stageID element. Usually used to indicate role, e.g. 'prod', 'staging', 'source', 'build', 'test', 'deploy', 'release'stringnullno
tagsAdditional tags (e.g. {'BusinessUnit': 'XYZ'}).
Neither the tag keys nor the tag values will be modified by this module.
map(string){}no
tenantID element _(Rarely used, not included by default)_. A customer identifier, indicating who this instance of a resource is forstringnullno
timeoutTime in seconds to wait for any individual kubernetes operation (like Jobs for hooks). Defaults to 300 secondsnumbernullno
waitWill wait until all resources are in a ready state before marking the release as successful. It will wait for as long as timeout. Defaults to trueboolnullno

Outputs

NameDescription
metadataBlock status of the deployed release

CHANGELOG

Version 1.445.0

Components PR #1039

Major Breaking Changes

Karpenter at version v0.33.0 transitioned from the v1alpha API to the v1beta API with many breaking changes. This component (eks/karpenter) changed as well, dropping support for the v1alpha API and adding support for the v1beta API. At the same time, the corresponding eks/karpenter-provisioner component was replaced with the eks/karpenter-node-pool component. The old components remain available under the deprecated/ directory.

The full list of changes in Karpenter is too extensive to repeat here. See the Karpenter v1beta Migration Guide and the Karpenter Upgrade Guide for details.

While a zero-downtime upgrade is possible, it is very complex and tedious and Cloud Posse does not support it at this time. Instead, we recommend you delete your existing Karpenter Provisioner (karpenter-provisioner) and Controller (karpenter) deployments, which will scale your cluster to zero and leave all your pods suspended, and then deploy the new components, which will resume your pods.

Full details of the recommended migration process for these components can be found in the Migration Guide.

If you require a zero-downtime upgrade, please contact Cloud Posse professional services for assistance.

Version 1.348.0

Components PR #868

The karpenter-crd helm chart can now be installed alongside the karpenter helm chart to automatically manage the lifecycle of Karpenter CRDs. However since this chart must be installed before the karpenter helm chart, the Kubernetes namespace must be available before either chart is deployed. Furthermore, this namespace should persist whether or not the karpenter-crd chart is deployed, so it should not be installed with that given helm-release resource. Therefore, we've moved namespace creation to a separate resource that runs before both charts. Terraform will handle that namespace state migration with the moved block.

There are several scenarios that may or may not require additional steps. Please review the following scenarios and follow the steps for your given requirements.

Upgrading an existing eks/karpenter deployment without changes

If you currently have eks/karpenter deployed to an EKS cluster and have upgraded to this version of the component, no changes are required. var.crd_chart_enabled will default to false.

Upgrading an existing eks/karpenter deployment and deploying the karpenter-crd chart

If you currently have eks/karpenter deployed to an EKS cluster, have upgraded to this version of the component, do not currently have the karpenter-crd chart installed, and want to now deploy the karpenter-crd helm chart, a few additional steps are required!

First, set var.crd_chart_enabled to true.

Next, update the installed Karpenter CRDs in order for Helm to automatically take over their management when the karpenter-crd chart is deployed. We have included a script to run that upgrade. Run the ./karpenter-crd-upgrade script or run the following commands on the given cluster before deploying the chart. Please note that this script or commands will only need to be run on first use of the CRD chart.

Before running the script, ensure that the kubectl context is set to the cluster where the karpenter helm chart is deployed. In Geodesic, you can usually do this with the set-cluster command, though your configuration may vary.

set-cluster <tenant>-<region>-<stage> terraform

Then run the script or commands:

kubectl label crd awsnodetemplates.karpenter.k8s.aws provisioners.karpenter.sh app.kubernetes.io/managed-by=Helm --overwrite
kubectl annotate crd awsnodetemplates.karpenter.k8s.aws provisioners.karpenter.sh meta.helm.sh/release-name=karpenter-crd --overwrite
kubectl annotate crd awsnodetemplates.karpenter.k8s.aws provisioners.karpenter.sh meta.helm.sh/release-namespace=karpenter --overwrite
info

Previously the karpenter-crd-upgrade script included deploying the karpenter-crd chart. Now that this chart is moved to Terraform, that helm deployment is no longer necessary.

For reference, the karpenter-crd chart can be installed with helm with the following:

helm upgrade --install karpenter-crd oci://public.ecr.aws/karpenter/karpenter-crd --version "$VERSION" --namespace karpenter

Now that the CRDs are upgraded, the component is ready to be applied. Apply the eks/karpenter component and then apply eks/karpenter-provisioner.

Note for upgrading Karpenter from before v0.27.3 to v0.27.3 or later

If you are upgrading Karpenter from before v0.27.3 to v0.27.3 or later, you may need to run the following command to remove an obsolete webhook:

kubectl delete mutatingwebhookconfigurations defaulting.webhook.karpenter.sh

See the Karpenter upgrade guide for more details.

Upgrading an existing eks/karpenter deployment where the karpenter-crd chart is already deployed

If you currently have eks/karpenter deployed to an EKS cluster, have upgraded to this version of the component, and already have the karpenter-crd chart installed, simply set var.crd_chart_enabled to true and redeploy Terraform to have Terraform manage the helm release for karpenter-crd.

Net new deployments

If you are initially deploying eks/karpenter, no changes are required, but we recommend installing the CRD chart. Set var.crd_chart_enabled to true and continue with deployment.