Why I Ditched the Cloud (For My Own)

My first attempt at running Cloudhaven was on Amazon EKS.
It was smooth, fast, and exactly like the setup I use at work — until the bill came in.
$280. For one month.

That was my “never again” moment.

I realized all I really needed were machines that can run Linux,
clustered together on the same network, and expose an API I could talk to.
Surely there had to be a cheaper (and more personal) way to do that.

The Birth of a Tiny Cloud ☁️

I already had a Raspberry Pi 4 (8GB RAM) lying around from an old project —
originally bought to host a Pi-hole to block ads across my network.

That little Pi became my first worker node.

I then bought an Orange Pi, which turned out to be a mistake —
limited community support, frustrating driver issues, and inconsistent docs.
So I returned it, picked up another Raspberry Pi, and paired it with an NVMe SSD kit (512GB).
That SSD-backed Pi became my control plane node.

At the time of writing, I’m running three nodes:

  1. Raspberry Pi (Master node, NVMe-backed)
  2. Raspberry Pi (Worker node)
  3. Mac mini VM (Worker node, via UTM)

All wired into the same switch — a tiny, humming, personal cloud in my living room.

The Stack: Talos Linux + Kubernetes

Instead of running full Linux distributions, I went with Talos Linux — a lightweight, immutable OS built specifically for Kubernetes.

Talos takes over the kernel, stripping away everything unnecessary, and exposes a secure API that you interact with using talosctl.

That means:

  • No SSH into nodes (everything is API-driven)
  • Immutable filesystem (great for security)
  • Minimal attack surface
  • Consistent upgrades and reboots

It’s frustrating at first (especially when debugging network configs), but the payoff in reliability and transparency is huge.

Keeping It Cheap — and Reliable

So far, the setup has been super efficient and surprisingly stable.
Each node idles around 2–4 watts, and everything boots in under 30 seconds.

To improve fault tolerance, here’s what I’ve added (or plan to add):

  • Daily Cluster Backups:
    Using Kubernetes’ batch API, I run a CronJob that exports etcd snapshots daily and uploads them to S3.
    (Still working on OIDC authentication for the service account — static keys would be easier, but I’m aiming for a production-grade setup.)

  • Uninterruptible Power Supply (UPS):
    Small UPS units can keep the control plane alive through short outages.
    Considering one for the switch + Pi master node combo.

  • Network Monitoring & Observability:
    I recently discovered HyperDX — an open-source, Datadog-adjacent observability stack built on ClickHouse and OpenTelemetry.
    I love that it relies on open standards (OTel especially), which means I can always pivot to another backend if I outgrow it.
    For now, I’m excited to integrate it into my cluster and see how well it handles logs, metrics, and traces in a lightweight homelab environment.

  • Future: Cluster Autoscaling with Pi Additions.
    The goal is to eventually expand to 4–5 nodes and introduce a small load balancer.

What I’ve Learned So Far

  • Cloud is convenient — but it’s not the only way.
    You can learn 90% of the same lessons locally for 1% of the cost.
  • Failure is cheap here.
    If my cluster goes down, I just reboot and keep going.
  • Talos Linux will test your patience.
    But it rewards you with a hardened, secure, API-first environment.
  • Building things from scratch builds empathy.
    I now understand Kubernetes control planes, networking, and storage in ways that EKS abstracts away.

Closing Thoughts

This project started with frustration — and turned into fascination.
I now have my own mini data center, completely under my control, running on open-source software and cheap hardware.

It’s not perfect, but it’s mine, and every crash, fix, and tweak teaches me something new.

Next, I’ll be exploring GitOps for my homelab — using ArgoCD to manage app deployments and experimenting with lightweight observability stacks like HyperDX.

Stay tuned — Cloudhaven’s local edition is just getting started.

🧠 Curious about Talos or my cluster layout? I’ve open-sourced part of my setup here:
https://github.com//cloudhaven/infrastructure