AWS Tf Conventions

AWS Terraform Conventions

Purpose

Apply these rules to all AWS foundation Terraform generated for KAS/NVCF EKS. This skill owns conventions only; use component skills for resource-specific content.

Terraform Ownership

Terraform owns these resources:

• VPC and subnets
• Internet gateway, NAT gateway, route tables, and subnet route associations
• EKS cluster and OIDC provider
• EKS managed node groups: system, compute, and gpu
• IAM roles and policies for the cluster and node groups
• ECR repositories under a stable cluster prefix, for example ${CLUSTER_NAME}/...
• Security groups needed for EKS-managed infrastructure only
• Terraform outputs needed by later skills

Terraform must not own:

• Kubernetes in-cluster objects such as namespaces, services, service accounts, RBAC, ConfigMaps, or storage classes
• Helm releases or Helmfile environments
• NVCF control plane, NVCA, LLS, caches, UCC, DDCS, storage APIs, functions, or stream validation
• NGC to ECR image copy jobs
• Secret values or registry auth artifacts
• GPU Operator Helm release state

Project Rules

1. Keep image mirroring separate from infra. Infra may create ECR repositories, but it must not copy NGC images as part of terraform apply.
2. Do not hardcode image versions in Terraform.
3. Do not create login material or machine access material with Terraform.
4. Do not use null_resource or local-exec for mirror jobs, waiters, or imperative Kubernetes work.
5. Use AWS-managed EKS behavior where possible. App ingress belongs to Kubernetes controllers/services after infra.
6. Use explicit per-environment CIDRs. Do not auto-select CIDRs unless the repo already has a reviewed helper for that.
7. Keep state safe. Prefer a remote backend for shared work; if local state is used, never run concurrent applies against the same env.
8. Add consistent tags to all taggable resources: Name, cluster, owner, environment, managed-by=terraform, and workload=kas.
9. Use the EKS GPU-optimized AMI for the GPU managed node group: ami_type = "AL2023_x86_64_NVIDIA".
10. Label GPU nodes with nvidia.com/gpu.deploy.driver=false; GPU Operator must not install the driver or container toolkit.
11. Keep Helm releases outside Terraform state.
12. Do not store secrets in *.tf, *.tfvars, generated scripts, or Terraform state when avoidable.

Expected Layout

Use a root stack per environment. A simple validation layout is acceptable:

terraform/aws/envs/<cluster>/
├── main.tf
├── providers.tf
├── variables.tf
├── outputs.tf
└── terraform.tfvars

When generating review artifacts, prefix the same layout with generated/:

generated/terraform/aws/envs/<cluster>/

For promoted code, split into modules only when it removes real complexity:

terraform/aws/modules/network
terraform/aws/modules/eks
terraform/aws/modules/ecr

Provider And Backend

Require explicit region input:

variable "aws_region" {
  type = string
}

provider "aws" {
  region = var.aws_region
}

Prefer a remote backend for shared environments. If using local state for a single-user validation environment, call that out in the generated README or run instructions and do not run concurrent applies.

Pre-Apply Checklist

• terraform fmt -recursive has run.
• terraform validate passes.
• terraform plan contains only expected AWS foundation resources.
• No plan action destroys or replaces a resource without explicit review.
• Tags are present on all taggable resources.
• No secrets or kubeconfig contents appear in Terraform files or outputs.
• GPU Operator, NVCF, and image mirroring are outside Terraform.