The AWS account Kestrel runs in: one VPC, one EKS cluster, two environments, and a CI pipeline that holds no AWS credentials.
Kestrel is a small European logistics company with one product and no
dedicated platform team. It needs an AWS account that is boring: two
environments that differ only in the ways they should, state that cannot be
corrupted by two people running apply at once, and a deploy pipeline that
does not require handing long-lived AWS keys to GitHub.
That last part is the interesting one. There is no AWS_ACCESS_KEY_ID in this
repository's secrets. GitHub mints a signed OIDC token that says which repo and
which branch is asking, AWS trusts that specific combination and nothing else,
and CI gets credentials that expire in an hour. There is no key to rotate, and
a leaked repository secret would be worth nothing because there isn't one.
flowchart TB
subgraph gh["GitHub"]
pr["Pull request"] -->|"plan, commented on the PR"| oidc
main["Merge to main"] -->|"apply, gated on an environment"| oidc
oidc(["OIDC token<br/>no stored credentials"])
end
oidc -->|"sts:AssumeRoleWithWebIdentity"| role["IAM role<br/>trusts only this repo + branch"]
subgraph aws["AWS account (eu-west-1)"]
role --> state[("S3 state bucket<br/>+ DynamoDB lock table")]
subgraph vpc["VPC 10.30.0.0/16"]
direction LR
subgraph az_a["eu-west-1a"]
pub_a["public /20<br/>NAT · ALB"]
priv_a["private /20<br/>EKS nodes"]
end
subgraph az_b["eu-west-1b"]
pub_b["public /20<br/>NAT · ALB"]
priv_b["private /20<br/>EKS nodes"]
end
end
role --> eks["EKS control plane<br/>access entries · audit logs"]
eks --> priv_a
eks --> priv_b
eks --> irsa["IRSA roles<br/>EBS CSI · ALB controller · external-dns"]
end
Everything except apply runs offline with no AWS account:
git clone https://github.com/thisiskazem/aws-landing-zone
cd aws-landing-zone
make check # fmt + validate + tflint across all 5 directoriesTo actually build it, apply bootstrap/ once by hand (it creates the state
backend and the CI role), put the bucket name into envs/*/backend.tf, then:
make plan ENV=dev
make apply ENV=dev # ~138 EUR/month if left running
make destroy ENV=dev # do this when you are doneWorth stating plainly, because "I forgot it was running" is the most common way a learning AWS account produces a surprising bill.
| dev | prod | |
|---|---|---|
| NAT gateways | 1 (~33 EUR) | 2 (~66 EUR) |
| EKS control plane | ~70 EUR | ~70 EUR |
| Nodes | 2 x t3.large Spot (~35 EUR) | 3 x m6i.large on-demand (~190 EUR) |
| Fixed monthly total | ~138 EUR | ~326 EUR |
make cost ENV=dev prints this from Terraform outputs. make destroy ENV=dev
takes it back to zero.
bootstrap/ S3 state bucket, DynamoDB lock table, GitHub OIDC provider and role
Applied once, by hand. Keeps its own state locally, by necessity.
modules/network/ VPC, public/private subnets across 2 AZs, NAT, flow logs
modules/platform/ EKS via terraform-aws-modules/eks v21, addons, three IRSA roles
envs/dev/ Spot nodes, one NAT, 7-day log retention
envs/prod/ On-demand nodes, NAT per AZ, 90-day retention
.github/workflows/ validate everywhere · plan on PR · apply gated on an environment
| Skill | Where to look |
|---|---|
| Terraform module design | modules/network/: derived CIDRs, validated inputs, no hardcoded values |
| Knowing what not to build | modules/platform/main.tf: the EKS control plane comes from upstream |
| AWS networking | modules/network/README.md: CIDR plan, subnet tags the ALB controller needs |
| IAM and least privilege | bootstrap/main.tf: OIDC trust policy scoped to named branches, IRSA per service account |
| EKS operations | access entries over aws-auth, IMDSv2 hop limit 1, prefix delegation, audit logs |
| GitHub Actions | .github/workflows/terraform.yaml: plan-on-PR, saved plan applied unchanged |
| Cost awareness | NAT topology as a variable, Spot in dev, cost in the outputs |
- The EKS module is not hand-written. Reimplementing 4000 lines of launch
templates and addon ordering would look impressive and be worse. The
hand-written module is
network/, where the decisions are actually mine. - The CI role has
AdministratorAccess. Terraform needs to create and destroy nearly anything, so a hand-scoped policy grows on every new resource type and gets a wildcard added during the first incident anyway. The control that carries the weight is the trust policy: only named branches of one named repository can assume it at all. - The API endpoint is public with a CIDR allowlist, not fully private. A
private cluster needs a bastion or VPN for every
kubectlcommand and for Argo CD to reach the API.0.0.0.0/0is rejected by a variable validation. - Two AZs, not three. The third adds a third NAT gateway and a third of everything to survive a failure mode two AZs already handle.
- State is encrypted with SSE-S3, not a customer-managed KMS key. KMS is
stronger and means every principal reading state also needs
kms:Decrypt. For a single-account setup that is complexity without a matching threat. - No RDS, ALB, WAF or CloudFront. They would each be twenty more files demonstrating nothing this repository does not already show.
The workloads that run on this cluster are delivered by gitops-k8s-platform.
Kazem Bigdeli · Cloud, DevOps & Platform Engineer · Turin, Italy linkedin.com/in/kazembigdeli