eks-nodes

---

name: eks-nodes description: Generate Terraform for additional EKS managed node groups: compute and GPU pools for KAS/NVCF workloads. Use after eks-cluster and before eks-storage. version: "1.0.0" author: NVIDIA Omniverse Streaming tags: ["aws", "eks", "node-groups", "gpu"] tools: ["Shell", "Read", "Write"]

EKS Additional Node Groups

What This Skill Produces

Generate Terraform for workload node groups on an existing EKS cluster:

• compute managed node group for CPU-heavy services such as streaming proxy or gateway-side helpers
• GPU managed node group for function workloads

Do not create the EKS control plane here. Do not install GPU Operator here.

Inputs

Required:

Input	Notes
CLUSTER_NAME	Existing EKS cluster name.
private_subnet_ids	Private subnet IDs from aws-iam-vpc.
node_role_arn or node IAM role config	Role used by managed node groups.

Defaults:

Input	Default
compute node instance type	m6i.2xlarge
compute min/max/desired	1 / 3 / 1
GPU node instance type	g6.12xlarge or closest quota-approved GPU SKU
GPU AMI type	AL2023_x86_64_NVIDIA
GPU min/max/desired	0 / 2 / 1

Node Group Labels

Required labels:

compute: node-type=compute
gpu:     node-type=gpu, nvidia.com/gpu.deploy.driver=false

Terraform module values should carry the same contract:

ami_type = "AL2023_x86_64_NVIDIA"

labels = {
  "node-type"                    = "gpu"
  "nvidia.com/gpu.deploy.driver" = "false"
}

The GPU node group must use ami_type = "AL2023_x86_64_NVIDIA" or a reviewed custom GPU AMI with the NVIDIA driver and container toolkit already installed.

Instance Metadata (IMDS)

Managed node groups launch instances with the default IMDS HttpPutResponseHopLimit = 1. A hop limit of 1 lets host-network pods reach IMDS but blocks ordinary (extra-hop) pods, so controllers that rely on the node instance-role credentials cannot read them. The aws-ebs-csi-driver controller (a normal Deployment) then CrashLoops with no EC2 IMDS role found … context deadline exceeded, while its hostNetwork node DaemonSet works — persistent-volume provisioning stays broken.

Each node group must therefore raise the hop limit to 2 via a launch template metadata_options block (so it survives scaling), and keep IMDSv2 required:

metadata_options {
  http_endpoint               = "enabled"
  http_tokens                 = "required"  # IMDSv2
  http_put_response_hop_limit = 2           # allow non-hostNetwork pods to reach IMDS
}

When the node group is created via aws_eks_node_group, attach this through a launch_template (managed node groups do not expose metadata_options directly). Alternatively, provision IRSA / EKS Pod Identity for the EBS CSI driver (and other controllers) so they do not depend on node-role IMDS at all; in that case the hop limit may stay at 1. Raising the hop limit is the simpler default and mirrors AWS best practice for EKS + IMDSv2.

Generation Rules

• Read aws-tf-conventions before generating Terraform.
• Keep GPU Operator out of Terraform; this skill only creates nodes.
• GPU desired_size defaults to 1 so a freshly applied cluster has a GPU node and eks-validate gpu-resource/cuda-smoke checks pass out of the box. If you intentionally set GPU min/desired to zero (scale-to-zero for cost), eks-validate will FAIL those checks unless you run it with ALLOW_GPU_ZERO=true, which downgrades them to WARN.
• Add taints only when the workload scheduling contract requires them; if a taint is used for GPU nodes, make sure validation and smoke-test pod include the matching toleration.

Validation Checklist

• Compute nodes are labeled node-type=compute.
• GPU nodes are labeled node-type=gpu.
• GPU nodes include nvidia.com/gpu.deploy.driver=false.
• GPU node group uses AL2023_x86_64_NVIDIA or reviewed equivalent.
• Autoscaling bounds are explicit.
• Node groups set IMDS http_put_response_hop_limit = 2 (via launch template) with http_tokens = "required", or IRSA / EKS Pod Identity is provisioned for IMDS-dependent controllers (e.g. aws-ebs-csi-driver).
• No Helm releases or Kubernetes objects are created here.