DEV Community

Arseny Zinchenko
Arseny Zinchenko

Posted on • Originally published at rtfm.co.ua on

Karpenter: the Beta version — an overview of changes, and upgrade from v0.30.0 to v0.32.1

#aws #devops #kubernetes

Karpenter: the Beta version — an overview of changes, and upgrade from v0.30.0 to v0.32.1

So, Karpenter has made another big step towards the release, and in version 0.32 it has moved from Alpha to Beta.

Let’s take a quick look at the changes — and they are quite significant — and then upgrade to EKS from Karpneter Terraform module and Karpenter Helm chart.

The process of installing Karpenter was described in the Terraform: building EKS, part 3 — Karpenter installation post, and below are some references to it, such as the file name karpenter.tf and some variables. Basic documentation:

What’s new in the v0.32.1?

Starting from 0.32, the Provisioner, AWSNodeTemplate and Machine API resources will be deprecated, and from version 0.33 they will be removed altogether.

They were replaced with:

  • Provisioner => NodePool
  • AWSNodeTemplate => EC2NodeClass
  • Machine => NodeClaim

v1alpha5/Provisioner => v1beta1/NodePool

NodePool is the successor to Provisioner, and has parameters for:

  • configuration of the Pods and WorkerNodes launch — requirements for Pods, tints, labels
  • configuration of how Karpenter will place Pods on Nodes and how it will perform the deprovisioning of unnecessary Nodes

“Under the hood”, NodePool will create v1beta1/NodeClaims resources (which replaced the v1alpha5/Machine resource), and in general, the idea here is similar to Deployments and Pods: in a Deployment (NodePool), we describe a template for how a Pod (NodeClaims) will be created. And v1beta1/NodeClaims, in turn, is the successor to the v1alpha5/Machine resource.

Also, a new section disruption has been added - all the parameters related to deleting redundant nodes and managing submissions have been moved here, see Disruption.

v1alpha5/AWSNodeTemplate => v1beta1/EC2NodeClass

EC2NodeClass is the successor to AWSNodeTemplate, and has parameters for:

  • AMI settings
  • SecurityGroups
  • subnets
  • EBS
  • IMDS
  • User Data

The spec.instanceProfile field has been removed - now Karpenter will create a Instance Profile for EC2 based on the IAM Role, which is passed to spec.role.

Also, the spec.launchTemplateName field has been finally removed.

See the documentation in NodeClasses.

Changes in the Labels

The karpenter.sh/do-not-evict and karpenter.sh/do-not-consolidate labels have been merged into a new label karpenter.sh/do-not-disrupt, which can be used both for Pods to prevent Karpenter from performing Pod evictions and for WorkerNodes to prevent the deletion of a Node.

Migration of 0.30.0 to v0.32.1

I’m going to describe below in some detail and consider different options, so it may seem like the upgrade process is quite a complicated because there’s a lot of text, but it’s not — it’s pretty simple.

What do we have now?

  • an AWS EKS Cluster created with Terraform
  • using the terraform-aws-modules/eks/aws//modules/karpenter module, the necessary resources are created in AWS - IAM, SQS, etc.
  • Karpenter itself is installed from the Helm-chart karpenter
  • and accepts a version which is passed from a var.helm_release_versions.karpenter variable
  • and we have two resources kubectl_manifest - for the karpenter_provisioner and karpenter_node_template

The migration process includes:

  1. update the IAM Role used by Karpenter controller views to manage EC2 in AWS:
  2. replace the karpenter.sh/provisioner-name tag with the karpenter.sh/nodepool (see chore: Release v0.32.0) - applies only to the role that was created from CloudFormation, because the Terraform module uses a different Condition
  3. add permissions to the IAM Policy — iam:CreateInstanceProfile, iam:AddRoleToInstanceProfile, iam:RemoveRoleFromInstanceProfile, iam:DeleteInstanceProfile, and iam:GetInstanceProfile
  4. adding new CRDs v1beta1, after which Karpenter will update Machine resources on NodeClaim by itself
  5. to migrate AWSNodeTemplate => EC2NodeClass and Provisioner => NodePool, you can use the karpenter-convert utility

Then replace the WorkerNodes, here we have three ways:

  • using the drift feature:
  • create a NodePool similar to the existing Provisioner
  • add a taint karpenter.sh/legacy=true:NoSchedule to the existing Provider
  • Karpenter will mark all Nodes of this Provider as drifted
  • Karpenter will launch new EC2s using new a NodePool
  • removal of nodes:
  • create a NodePool similar to the existing Provisioner
  • delete an existing Provisioner with the command kubectl delete provisioner <provisioner-name> --cascade=foreground, as a result of which Karpenter will delete all its Nodes by performing Node Drain for all at once, and Pods will go into the Pending state and will be scheduled on Nodes that were created from NodePool
  • manual replacement:
  • create a NodePool similar to the existing Provisioner
  • add a karpenter.sh/legacy=true:NoSchedule to the old Provider
  • manually delete all its WorkerNodes one by one with the kubectl delete node

What does this mean for us?

All the changes are seemingly backward compatible (I checked — rolled back versions), that is, we can safely update existing resources one by one — nothing should break.

So, what will we do?

  • update the Terraform module
  • add CRDs
  • update Karpenter’s Helm chart
  • deploy it, check it — old Nodes from the old Provisioner will continue to work until we kill them
  • add a new NodePool and EC2NodeClass
  • recreate the WorkerNodes

Let’s go.

Step 1: update the Terraform Karpenter module

First, run terraform apply before making any changes to have the latest version of your code deployed.

In the karpenter.tf file, we have the module call and its version:

module "karpenter" {
  source = "terraform-aws-modules/eks/aws//modules/karpenter"
  version = "19.16.0"
...
Enter fullscreen mode Exit fullscreen mode

Karpenter Instance Profile

In the module itself, a new variable enable_karpenter_instance_profile_creation has been added, which determines who will manage IAM Roles for WorkerNodes - Terraform, as it was before, or use a new feature from Karpenter. If enable_karpenter_instance_profile_creation is set to true, the module simply adds another block of rights to IAM, see main.tf.

But here "there is a nuance" (c) in the dependencies of the Karpenter module and the Helm chart: if enable_karpenter_instance_profile_creation is set to true (the default value is false), the module will not create resource "aws_iam_instance_profile", which is further used in the chart for Karpenter - the parameter settings.aws.defaultInstanceProfile.

So there are two options here:

  1. first update only the module and chart versions, but use the old parameters and Provisioner
  2. after the update — create a NodePool and EC2NodeClass, replace the parameters, and recreate WorkerNodes
  3. or update everything at once — the module/chart, parameters, and Provisioner to replace with NodePool, and seal everything together

At first, you can do it step by step, the option 1, somewhere on the Dev cluster, to see how it goes, and then roll out the entire update to Prod at once.

Let’s start, of course, with Dev.

Change the version to v19.18.0 — see Releases, and add enable_karpenter_instance_profile_creation = true - but for now, let's leave it commented with #:

module "karpenter" {
  source = "terraform-aws-modules/eks/aws//modules/karpenter"
  #version = "19.16.0"
  version = "19.18.0"

  cluster_name = module.eks.cluster_name

  irsa_oidc_provider_arn = module.eks.oidc_provider_arn
  irsa_namespace_service_accounts = ["karpenter:karpenter"]

  create_iam_role = false
  iam_role_arn = module.eks.eks_managed_node_groups["${local.env_name_short}-default"].iam_role_arn
  irsa_use_name_prefix = false

  # In v0.32.0/v1beta1, Karpenter now creates the IAM instance profile
  # so we disable the Terraform creation and add the necessary permissions for Karpenter IRSA
  #enable_karpenter_instance_profile_creation = true
}

...
Enter fullscreen mode Exit fullscreen mode

Do not deploy it yet, let’s go to the chart.

Step 2: update the Karpenter Helm chart

I have the versions of the charts in one variable — update here karpneter v0.30.0 to v0.32.1:

...
helm_release_versions = {
  #karpenter = "v0.30.0"
  karpenter = "v0.32.1"
  external_dns = "1.13.1"
  aws_load_balancer_controller = "1.6.1"
  secrets_store_csi_driver = "1.3.4"
  secrets_store_csi_driver_provider_aws = "0.3.4"
  vpa = "2.5.1"
}
...
Enter fullscreen mode Exit fullscreen mode

Next, we have two major updates: CRD and chart values.

Karpenter CRD Upgrade

CRDs are in the main Karpenter chart, but:

Helm does not manage the lifecycle of CRDs using this method, the tool will only install the CRD during the first installation of the helm chart. Subsequent chart upgrades will not add or remove CRDs, even if the CRDs have changed. When CRDs are changed, we will make a note in the version’s upgrade guide.

That is, subsequent runs of helm install will not update the CRDs that were installed during the first installation.

So there are two options here (again!) - either just manually add them from kubectl, or install them from an additional chart karpenter-crd, see CRD Upgrades.

Moreover, the chart will install both the old CRDs v1alpha5 and the new ones v1beta1, that is, we will have a kind of "backward compatible mode" - we will be able to use the old Provisioner and add a new NodePool at the same time.

Using the helm template, you can check what exactly the karpenter-crd chart will do:

$ helm template karpenter-crd oci://public.ecr.aws/karpenter/karpenter-crd --version v0.32.1
...
Enter fullscreen mode Exit fullscreen mode

But there is a caveat here: to install new CRDs from the chart, you will need to delete existing CRDs, and this will result in the removal of the existing Provisioner and Machine and the corresponding WorkerNodes.

So if this is a long-existing cluster and you want to do everything without downtime, then new CRDs needs to be installed manually.

If a few minutes of downtime is okay for you, then it’s better to use an additional Helm chart, and then everything will continue to be managed automatically.

So, in my case, we update the CRDs with a chart.

Add it to the Terraform code:

...
resource "helm_release" "karpenter_crd" {
  namespace = "karpenter"
  create_namespace = true

  name = "karpenter-crd"
  repository = "oci://public.ecr.aws/karpenter"
  repository_username = data.aws_ecrpublic_authorization_token.token.user_name
  repository_password = data.aws_ecrpublic_authorization_token.token.password
  chart = "karpenter-crd"
  version = var.helm_release_versions.karpenter
}
...
Enter fullscreen mode Exit fullscreen mode

Let’s move on to the main chart.

Karpenter Chart values

Next, in the resource "helm_release" "karpenter" we write new values and add depends_on to the chart with CRDs.

Add the settings.aws.defaultInstanceProfile parameter - later, we will remove it:

...
resource "helm_release" "karpenter" {
  namespace = "karpenter"
  create_namespace = true

  name = "karpenter"
  repository = "oci://public.ecr.aws/karpenter"
  repository_username = data.aws_ecrpublic_authorization_token.token.user_name
  repository_password = data.aws_ecrpublic_authorization_token.token.password
  chart = "karpenter"
  version = var.helm_release_versions.karpenter

  values = [
    <<-EOT
    settings:
      clusterName: ${module.eks.cluster_name}
      clusterEndpoint: ${module.eks.cluster_endpoint}
      interruptionQueueName: ${module.karpenter.queue_name}
      aws:
        defaultInstanceProfile: ${module.karpenter.instance_profile_name}
    serviceAccount:
      annotations:
        eks.amazonaws.com/role-arn: ${module.karpenter.irsa_arn} 
    EOT
  ]

  depends_on = [
    helm_release.karpenter_crd
  ]

  /*
  set {
    name = "settings.aws.clusterName"
    value = local.env_name
  }

  set {
    name = "settings.aws.clusterEndpoint"
    value = module.eks.cluster_endpoint
  }

  set {
    name = "serviceAccount.annotations.eks\\.amazonaws\\.com/role-arn"
    value = module.karpenter.irsa_arn
  }

  set {
    name = "serviceAccount.annotations.eks\\.amazonaws\\.com/sts-regional-endpoints"
    value = "true"
    type = "string"
  }

  set {
    name = "settings.aws.defaultInstanceProfile"
    value = module.karpenter.instance_profile_name
  }

  set {
    name = "settings.aws.interruptionQueueName"
    value = module.karpenter.queue_name
  }
  */
}
...
Enter fullscreen mode Exit fullscreen mode

Execute terraform init to load the new version of the module.

Now, we already have CRDs that were created during the first installation of Karpenter:

$ kk get crd | grep karpenter
awsnodetemplates.karpenter.k8s.aws 2023-10-03T08:30:58Z
machines.karpenter.sh 2023-10-03T08:30:59Z
provisioners.karpenter.sh 2023-10-03T08:30:59Z
Enter fullscreen mode Exit fullscreen mode

Delete them ( this will lead to the Provisioners, Machines, and Karpenter’s WorkerNodes removal! ):

$ kk -n karpenter delete crd awsnodetemplates.karpenter.k8s.aws machines.karpenter.sh provisioners.karpenter.sh
customresourcedefinition.apiextensions.k8s.io "awsnodetemplates.karpenter.k8s.aws" deleted
customresourcedefinition.apiextensions.k8s.io "machines.karpenter.sh" deleted
customresourcedefinition.apiextensions.k8s.io "provisioners.karpenter.sh" deleted
Enter fullscreen mode Exit fullscreen mode

And run terraform apply - now only the chart itself should be updated - there will be no changes in IAM yet, because we have enable_karpenter_instance_profile_creation == false.

After deployment, check the CRDs:

$ kk get crd | grep karpenter
awsnodetemplates.karpenter.k8s.aws 2023-11-02T15:33:26Z
ec2nodeclasses.karpenter.k8s.aws 2023-11-03T11:20:07Z
machines.karpenter.sh 2023-11-02T15:33:26Z
nodeclaims.karpenter.sh 2023-11-03T11:20:08Z
nodepools.karpenter.sh 2023-11-03T11:20:08Z
provisioners.karpenter.sh 2023-11-02T15:33:26Z
Enter fullscreen mode Exit fullscreen mode

Check your Pods and Nodes — everything should remain as it was — the same Machine resource that was created from the old Provisioner, and the same WorkerNode:

$ kk get machine
NAME TYPE ZONE NODE READY AGE
default-b6hdr t3.large us-east-1a ip-10-1-35-97.ec2.internal True 17m
Enter fullscreen mode Exit fullscreen mode

If everything is OK, then proceed to creating a NodePool and EC2NodeClass.

Step 3: Create a NodePool and EC2NodeClass

First, let’s look at the IAM roles :-) But this applies to my current setup, so if you create all your nodes with Karpenter, you can skip this part.

In the EKS Terraform module, we create a Managed Node Group, in which we create an IAM Role, which is then used in the InstanceProfile for all nodes in the cluster.

Then this Role is passed to the karpenter module, and therefore create_iam_role in the Karpenter module is set to false - because the role already exists:

...
module "karpenter" {
  ...
  # disable create as doing in EKS NodeGroup resource
  create_iam_role = false
  iam_role_arn = module.eks.eks_managed_node_groups["${local.env_name_short}-default"].iam_role_arn
  irsa_use_name_prefix = false
  ...
}
...
Enter fullscreen mode Exit fullscreen mode

Then, when Karpenter launched new EC2 instances, they were assigned this Role.

But with the new version of Karpenter, he creates its own instanceProfile from the spec.role.

To pass the name instead of iam_role_arn in the new manifest from EC2NodeClass, we look for it in the outputs.tf:

...
output "iam_role_name" {
  description = "The name of the IAM role"
  value = try(aws_iam_role.this[0].name, null)
}

output "iam_role_arn" {
  description = "The Amazon Resource Name (ARN) specifying the IAM role"
  value = try(aws_iam_role.this[0].arn, var.iam_role_arn)
}
...
Enter fullscreen mode Exit fullscreen mode

Now you can add the rest of the resources.

Adding an EC2NodeClass

See the documentation in the NodeClasses.

Here, we do it directly in the code of the file karpenter.tf, as it was for the AWSNodeTemplate.

We keep the old manifest and add a new one next to it:

...
resource "kubectl_manifest" "karpenter_node_template" {
  yaml_body = <<-YAML
    apiVersion: karpenter.k8s.aws/v1alpha1
    kind: AWSNodeTemplate
    metadata:
      name: default
    spec:
      subnetSelector:
        karpenter.sh/discovery: "atlas-vpc-${var.environment}-private"
      securityGroupSelector:
        karpenter.sh/discovery: ${local.env_name}
      tags:
        Name: ${local.env_name_short}-karpenter
        environment: ${var.environment}
        created-by: "karpneter"
        karpenter.sh/discovery: ${local.env_name}
  YAML

  depends_on = [
    helm_release.karpenter
  ]
}

resource "kubectl_manifest" "karpenter_node_class" {
  yaml_body = <<-YAML
    apiVersion: karpenter.k8s.aws/v1beta1
    kind: EC2NodeClass
    metadata:
      name: default
    spec:
      amiFamily: AL2
      role: ${module.eks.eks_managed_node_groups["${local.env_name_short}-default"].iam_role_name}
      subnetSelectorTerms:
        - tags:
            karpenter.sh/discovery: "atlas-vpc-${var.environment}-private"
      securityGroupSelectorTerms:
        - tags:
            karpenter.sh/discovery: ${local.env_name}
      tags:
        Name: ${local.env_name_short}-karpenter
        environment: ${var.environment}
        created-by: "karpneter"
        karpenter.sh/discovery: ${local.env_name}
  YAML

  depends_on = [
    helm_release.karpenter
  ]
}
...
Enter fullscreen mode Exit fullscreen mode

In spec.amiFamily we pass the Amazon Linux v2, in the spec.role - an IAM Role for the InstanceProfile - it is also added to the aws-auth ConfigMap of our cluster (in the eks module).

Adding a NodePool

Since it was planned to have more than one Provisioner/NodePools, their parameters are set in variables — copy the variable itself:

...
variable "karpenter_provisioner" {
  type = map(object({
    instance-family = list(string)
    instance-size = list(string)
    topology = list(string)
    labels = optional(map(string))
    taints = optional(object({
      key = string
      value = string
      effect = string
    }))
  }))
}

variable "karpenter_nodepool" {
  type = map(object({
    instance-family = list(string)
    instance-size = list(string)
    topology = list(string)
    labels = optional(map(string))
    taints = optional(object({
      key = string
      value = string
      effect = string
    }))
  }))
}
...
Enter fullscreen mode Exit fullscreen mode

And the values:

...
karpenter_provisioner = {
  default = {
    instance-family = ["t3"]
    instance-size = ["small", "medium", "large"]
    topology = ["us-east-1a", "us-east-1b"]
    labels = {
      created-by = "karpenter"
    }
  }
}

karpenter_nodepool = {
  default = {
    instance-family = ["t3"]
    instance-size = ["small", "medium", "large"]
    topology = ["us-east-1a", "us-east-1b"]
    labels = {
      created-by = "karpenter"
    }
  }
}
...
Enter fullscreen mode Exit fullscreen mode

Just like with Provisioner, we add the template file configs/karpenter-nodepool.yaml.tmpl - here the format has also changed a bit, for example, labels is now in the spec.template.metadata.labels block instead of spec.labels as it was in Provisioner, see NodePools.

So now the template look like this:

apiVersion: karpenter.sh/v1beta1
kind: NodePool
metadata:
  name: ${name}

spec:
  template:

    metadata:

    %{ if labels != null ~} 
      labels:
      %{ for k, v in labels ~}
        ${k}: ${v}
      %{ endfor ~}    
    %{ endif ~}

    spec:

    %{ if taints != null ~}
      taints:
        - key: ${taints.key}
          value: ${taints.value}
          effect: ${taints.effect}
    %{ endif ~}

      nodeClassRef:
        name: default
      requirements:
        - key: karpenter.k8s.aws/instance-family
          operator: In
          values: ${jsonencode(instance-family)}
        - key: karpenter.k8s.aws/instance-size
          operator: In
          values: ${jsonencode(instance-size)}
        - key: topology.kubernetes.io/zone
          operator: In
          values: ${jsonencode(topology)}
  # total cluster limits 
  limits:
    cpu: 1000
    memory: 1000Gi
  disruption:
    consolidationPolicy: WhenEmpty
    consolidateAfter: 30s
Enter fullscreen mode Exit fullscreen mode

Important : if you are not using the Spot Instances in AWS, then add the karpenter.sh/capacity-type == "on-demand", see the reason below in the Error: The provided credentials do not have permission to create the service-linked role for EC2 Spot Instances.

And add a new kubectl_manifest resource to karpenter.tf, next to the old Provisioner:

...
resource "kubectl_manifest" "karpenter_provisioner" {
  for_each = var.karpenter_provisioner

  yaml_body = templatefile("${path.module}/configs/karpenter-provisioner.yaml.tmpl", {
    name = each.key
    instance-family = each.value.instance-family
    instance-size = each.value.instance-size
    topology = each.value.topology
    taints = each.value.taints
    labels = merge(
      each.value.labels,
      {
        component = var.component
        environment = var.environment
      }
    )
  })

  depends_on = [
    helm_release.karpenter
  ]
}

resource "kubectl_manifest" "karpenter_nodepool" {
  for_each = var.karpenter_nodepool

  yaml_body = templatefile("${path.module}/configs/karpenter-nodepool.yaml.tmpl", {
    name = each.key
    instance-family = each.value.instance-family
    instance-size = each.value.instance-size
    topology = each.value.topology
    taints = each.value.taints
    labels = merge(
      each.value.labels,
      {
        component = var.component
        environment = var.environment
      }
    )
  })

  depends_on = [
    helm_release.karpenter
  ]
}
...
Enter fullscreen mode Exit fullscreen mode

Next:

  • in the resource "helm_release" "karpenter" remove the aws.defaultInstanceProfile from the values
  • in the module "karpenter" enable the enable_karpenter_instance_profile_creation option

Now Terraform will:

  • add permissions to the KarpenterIRSA IAM Role
  • if the create_instance_profile == false parameter was not passed for the karpenter module, then module.karpenter.aws_iam_instance_profile will be deleted, but in my case it was not used anyway
  • add kubectl_manifest.karpenter_nodepool["default"] and kubectl_manifest.karpenter_node_class resources

Deploy it and check it:

$ kk get nodepool
NAME NODECLASS
default default

$ kk get ec2nodeclass
NAME AGE
default 40s
Enter fullscreen mode Exit fullscreen mode

And we still have our old Machine:

$ kk get machine
NAME TYPE ZONE NODE READY AGE
default-b6hdr t3.large us-east-1a ip-10-1-35-97.ec2.internal True 33m
Enter fullscreen mode Exit fullscreen mode

That’s it — we just need to recreate WorkerNodes, move Pods, and after deploying to Staging and Production, clean up the code — remove everything that’s left of version 0.30.0.

Error: “The provided credentials do not have permission to create the service-linked role for EC2 Spot Instances”

At some point, an error happened in the logs:

karpenter-5dcf76df9-l58zq:controller {“level”:”ERROR”,”time”:”2023–11–03T14:59:41.072Z”,”logger”:”controller”,”message”:”Reconciler error”,”commit”:”1072d3b”,”controller”:”nodeclaim.lifecycle”,”controllerGroup”:”karpenter.sh”,”controllerKind”:”NodeClaim”,”NodeClaim”:{“name”:”default-ttx86"},”namespace”:””,”name”:”default-ttx86",”reconcileID”:”6d17cadf-a6ca-47e3–9789–2c3491bf419f”,”error”:”launching nodeclaim, creating instance, with fleet error(s), AuthFailure.ServiceLinkedRoleCreationNotPermitted: The provided credentials do not have permission to create the service-linked role for EC2 Spot Instances.”}

But first, why Spot? Where did it come from?

If you look at the NodeClaim that was created for this Node, you will see "karpenter.sh/capacity-type == spot":

$ kk get nodeclaim -o yaml
...
  spec:
    ...
    - key: karpenter.sh/nodepool
      operator: In
      values:
      - default
    - key: karpenter.sh/capacity-type
      operator: In
      values:
      - spot
...
Enter fullscreen mode Exit fullscreen mode

Although the documentation says that the default capacity-type should be on-demand:

...
        - key: "karpenter.sh/capacity-type" # If not included, the webhook for the AWS cloud provider will default to on-demand
          operator: In
          values: ["spot", "on-demand"]
...
Enter fullscreen mode Exit fullscreen mode

And we didn’t specify it in NodePool, so must be the default. If you specify karpenter.sh/capacity-type explicitly in NodePool:

...
      requirements:
        - key: karpenter.k8s.aws/instance-family
          operator: In
          values: ${jsonencode(instance-family)}
        - key: karpenter.k8s.aws/instance-size
          operator: In
          values: ${jsonencode(instance-size)}
        - key: topology.kubernetes.io/zone
          operator: In
          values: ${jsonencode(topology)}
        - key: karpenter.sh/capacity-type
          operator: In 
          values: ["on-demand"]
...
Enter fullscreen mode Exit fullscreen mode

Then everything is working as it should.

And secondly, what permissions is it missing? What is the “ServiceLinkedRoleCreationNotPermitted" error?

I've never used Spots in AWS, so had to do some googling, and the answer was found in the documentation Work with Spot Instances and the guide Using AWS Spot instances, which says about the AWSServiceRoleForEC2Spot IAM Role that must be created in the AWS Account to be able to create Spot instances.

It's a bit weird to create Spots by default, especially since the documentation says otherwise. In addition, in 0.30, everything worked without explicitly setting karpenter.sh/capacity-type.

Okay, let's keep in mind that if you use On Demand exclusively, you need to add NodePool to the configuration.

Step 3: update WorkerNodes

All we have to do is move our Pods to new Nodes.

In fact, all Pods were moved to new Nodes during the update, but let’s do it, because we still have other clusters to update.

Here we have three options, which we discussed at the beginning. Let’s try to do it without downtime — using drift (but without downtime means if you have at least 2 Pods per service, and in addition PodDisruptionBudget).

What we need to do is add a taint to the existing Provisioner, deploy the changes so that taint is added to Nodes, and then Karpenter will perform Node Drain and create new Nodes to move our workloads.

Update the configs/karpenter-provisioner.yaml.tmpl with the taint:

apiVersion: karpenter.sh/v1alpha5
kind: Provisioner
metadata:
  name: ${name}
spec:

  taints:
    - key: karpenter.sh/legacy
      value: "true"
      effect: NoSchedule
...
Enter fullscreen mode Exit fullscreen mode

In the 0.32.1 chart, the drift parameter is still false, so we include featureGates.drift to our resource "helm_release" "karpenter":

...
  values = [
    <<-EOT
    settings:
      clusterName: ${module.eks.cluster_name}
      clusterEndpoint: ${module.eks.cluster_endpoint}
      interruptionQueueName: ${module.karpenter.queue_name}
      featureGates:
        drift: true
...
Enter fullscreen mode Exit fullscreen mode

And all these changes can be rolled out to other Kubernetes clusters together — just don’t forget to update the tfvars for these clusters (if you have something like separate envs/dev/dev-1-28.tfvars, envs/staging/staging-1-28.tfvars, envs/prod/prod-1-28.tfvars).

Rolling back the upgrade

It’s unlikely that this will be necessary, because there shouldn’t be any problems, but I did it during the re-test of the upgrade, so I’ll write it down:

  • change the version of module "karpenter" from the new 19.18.0 to the old 19.16.0
  • in module "karpenter" comment out the option enable_karpenter_instance_profile_creation
  • in the tfvars for the helm_release_versions change the version of the Karpenter chart from the new v0.32.1 to the old v0.30.0
  • leave the resource "helm_release" "karpenter_crd"
  • in resource "helm_release" "karpenter" comment out the new block of values, uncomment the old values (with the set)
  • comment out the resources resource "kubectl_manifest" "karpenter_nodepool" and resource "kubectl_manifest" "karpenter_node_class"
  • in the file configs/karpenter-provisioner.yaml.tmpl remove Taint

Originally published at RTFM: Linux, DevOps, and system administration.

Top comments (0)

Read next

danial profile image

Effective Strategies for Application Monitoring in AWS

Danial Ranjha -

hana_sato profile image

Talend vs. Apache Kafka: Which Data Tool Drives Better Business Insights?

Hana Sato -

vellanki profile image

How to Master Kubernetes Troubleshooting? Do it with 35 Real-World Scenarios

Koti Vellanki -

094459 profile image

Amazon Q Developer Tips: No.17 Choose the right tool

Ricardo Sueiras -