A Free Kubernetes Cluster on Oracle Cloud's Ampere Free Tier
Four Ampere arm64 nodes, three control planes spread across fault domains, Longhorn storage, backups to B2 and a live Ghost site, all on Oracle Cloud's Always Free tier for $0.00 a month.
My homelab cluster runs in a closet. That's great until the power goes out, the internet goes down, or I'm in the middle of an upgrade. I wanted a second cluster somewhere else: something that keeps running when home doesn't, and somewhere to try things without risking the cluster everything else depends on.
Oracle Cloud's Always Free tier turned out to be a really good fit. This post covers what I built, what it's running, and why I think it's one of the best free ways to learn Kubernetes properly.
What you get for free
The interesting part of OCI's Always Free tier is the Ampere A1 allowance:
- 4 OCPUs of Arm (Ampere Altra) compute
- 24 GB of RAM
- 200 GB of block storage
You can carve that up however you like: one big VM, or several small ones. On Ampere, one OCPU is a full physical core rather than a hyperthread, so these aren't toy VMs. For a Kubernetes cluster, four separate nodes is the obvious choice.
The cluster: oci-talos1
Like the homelab, it runs Talos Linux: an immutable, API-driven OS with no SSH and no shell, built only to run Kubernetes. It's on Talos v1.14.5 and Kubernetes 1.37.1, one minor version ahead of home, which makes it a handy place to try a new Kubernetes release before the homelab gets it.
| Node | Role | Shape | CPU / RAM | Fault domain |
|---|---|---|---|---|
| oci-cp1 | control plane | VM.Standard.A1.Flex | 1 OCPU / 6 GB | FD-1 |
| oci-cp2 | control plane | VM.Standard.A1.Flex | 1 OCPU / 6 GB | FD-2 |
| oci-cp3 | control plane | VM.Standard.A1.Flex | 1 OCPU / 6 GB | FD-3 |
| oci-w1 | worker | VM.Standard.A1.Flex | 1 OCPU / 6 GB | FD-1 |
Each node also gets a ~47 GB boot volume, and workloads are allowed on the control planes. Four nodes use the whole free A1 allowance. A fifth would show up on the bill.
I could have run one control plane and three workers for a bit more usable capacity, but three control planes means a real etcd quorum. I can lose a control plane and the cluster carries on. Since the point is to learn how a production-style cluster behaves, I wanted that.
Spreading across fault domains
This is one of the best parts as a learning environment. OCI regions are split into availability domains (separate data centers), and each availability domain is split into three fault domains: separate racks with their own hardware and power.
The free tier's block storage only applies in your home region, which for me is San Jose, and San Jose has a single availability domain. So the control planes are pinned one per fault domain. Losing a rack, a host or a power feed takes out at most one control plane.
OCI doesn't do this for you. When I first launched the nodes, all four landed in the same fault domain. The launch script now reads each node's fault domain from the Ansible inventory and pins it explicitly. On a homelab, "different racks" isn't something you get to practice. Here it's free, and you can actually see topology-aware scheduling and failure domains work.
How it's built
Everything lives in its own Ansible repo, separate from the homelab on purpose. Nothing stretches etcd or storage across the internet between the two.
- Talos image: built with the Talos Image Factory with
iscsi-toolsandutil-linux-toolsadded for Longhorn. Same schematic as home, because both clusters are arm64. - No bootstrap dance: each VM boots with its Talos machine config passed in as OCI
user_data, so the usualapply-config --insecurestep is gone. A launch script creates the VM, and a playbook bootstraps etcd and fetches the kubeconfig. - Locked-down network: only one control plane has a public IP, and the security list only allows the Talos API and Kubernetes API from my home IP. The other nodes are on a private subnet.
talosctltalks to the public node, which proxies to the rest. - No VIP needed: there's no layer 2 network in the cloud for a floating IP, so the nodes reach the API through Talos's built-in KubePrism load balancer. Losing cp1 only blocks new nodes joining and my remote access. Running workloads don't notice.
Storage
Each node's boot disk is split in the Talos config: 20 GB for the OS (images, logs, etcd) and ~24.4 GB for a Longhorn volume at /var/mnt/longhorn. Longhorn runs with 2 replicas across nodes, which gives roughly 50 GB of replicated, redundant storage for $0.
Longhorn backs up to Backblaze B2 every night, under its own prefix, with a week of retention. Plus a nightly Longhorn system backup.
Getting traffic in
Same pattern as home: no inbound ports are open for websites. Traffic comes in through a Cloudflare Tunnel (three cloudflared replicas spread across the nodes) into Traefik using the Kubernetes Gateway API. The Cloudflare Zero Trust operator watches my HTTPRoutes and creates the tunnel routes and DNS records automatically. When I add a hostname to an HTTPRoute, it goes live.
What's running on it: www.dougiesmash.org
The first real workload is www.dougiesmash.org. It's a Ghost site backed by MySQL, both on Longhorn volumes, with its own custom theme. It used to be on a paid managed Ghost host. Now it runs on a cluster that costs nothing.
It lives on OCI instead of at home because it's the kind of site that should stay up when the homelab is down for maintenance or the house loses power.
How much is actually in use
Numbers from this week, with everything running: Talos, etcd on three nodes, Longhorn, Traefik, cloudflared, the operator, metrics-server, Ghost and MySQL.
| Node | CPU used | Memory used | CPU requested |
|---|---|---|---|
| oci-cp1 | ~160m (16%) | ~2.0 GB (38%) | 53% |
| oci-cp2 | ~170m (17%) | ~2.4 GB (46%) | 58% |
| oci-cp3 | ~200m (20%) | ~2.3 GB (44%) | 79% |
| oci-w1 | ~90m (9%) | ~1.2 GB (21%) | 47% |
Real usage is low: under a quarter of the CPU and well under half the memory. The tighter limit is requested CPU. Kubernetes reserves what pods ask for, and the control plane components and Longhorn's instance managers ask for a lot on a 1-core node. I turned Longhorn's instance manager CPU reservation down to 5% to make room.
Storage: Longhorn has about 30 GB scheduled (the Ghost content and MySQL volumes, each replicated twice) out of ~98 GB raw.
And the bill: $0.00.
The Ampere VMs are pretty decent
I'll be honest, I expected free-tier VMs to feel like free-tier VMs. They don't. Ghost is quick, the control plane is responsive, and pulling and starting images doesn't drag. A full Ampere core per OCPU makes a real difference compared with the shared, burstable instances most free tiers give you.
Being arm64 is a bonus for me. My homelab is Turing RK1 boards, also arm64, so every image I build in GitHub Actions runs on both clusters unchanged. If you're on x86 at home, it's a good push to start building multi-arch images.
Why this is a great way to learn Kubernetes
- It's a real multi-node cluster. Not minikube, not kind on a laptop. Real VMs, real networking, real etcd quorum.
- Failure domains you can actually use. Spreading across fault domains, pod topology spread, and what happens when a node disappears.
- Real cloud networking. VCNs, public and private subnets, security lists, and working without layer 2 tricks like VIPs.
- Persistent storage done properly. Distributed block storage with replicas and off-site backups, not just hostPath.
- It's always on. Run something public on it and you have to care about uptime, upgrades and backups, which is where most of the learning happens.
- It costs nothing. Break it, rebuild it from the repo, repeat.
A few caveats
- Capacity can be hard to get. A1 instances are popular and some regions report "out of capacity." OCI has a capacity report API you can check before launching.
- Upgrade to Pay As You Go. You still pay nothing if you stay inside the Always Free limits, but Oracle can reclaim idle Always Free instances on free-only accounts. Put a budget alert on it.
- Stay inside the lines. I'm at 4/4 OCPU, 24/24 GB and 188/200 GB of block storage. Any extra volume or VM starts costing money.
- Free can end. Oracle can change the free tier whenever it wants, so I treat this cluster as something I can rebuild. It's all in Ansible and the data goes to B2, so if OCI ever went away, dougiesmash.org could be restored onto the home cluster.
If you've wanted a real Kubernetes cluster to learn on but didn't want to buy hardware or pay a cloud bill, this is the one I'd point you to. Four Ampere cores and 24 GB of RAM goes a lot further than you'd think.