Skip to content

Cloud-at-home on mini-PCs

A seven-node heterogeneous Kubernetes cluster assembled, provisioned, and operated from Git. It is the foundation every other project runs on.

Platform & Operations Running datahub-local-bootstrap datahub-local-core

The DataHub.local cluster, a seven-node mini-PC rack in a dark studio setting
The seven-node cluster, from the UPS and network switch to the ARM boards and the GPU node

Problem

Running production-like workloads at home means starting without the safety nets a cloud provides: no managed control plane, no elastic capacity, and no vendor who owns the failure. The first attempt was a cluster of consumer ARM single-board computers. It worked for lightweight services but proved unreliable and slow for data workloads, with SD-card failures and firmware quirks that consumed more time than they saved.

The challenge was to turn a mixed pile of affordable hardware into something that behaves like a real platform.

Approach

The cluster is treated as a product with a delivery pipeline, not a collection of machines.

  • Heterogeneous by design: ARM64 boards carry lightweight and multi-arch workloads, while x86 mini-PCs, including a dedicated NAS and a GPU worker, handle Spark, Trino, and inference.
  • Provisioned as code: Ansible installs the OS and K3s and seeds ArgoCD. After that first run, Git is the only interface to the cluster.
  • Delivered via GitOps: ArgoCD reconciles one Application per namespace, layered as secrets, then core, then AI.
  • Storage matched to data: replicated block storage for databases, object storage for the lakehouse, and bulk RAID for media and archives.
  • Hardware chosen on evidence: the move to x86 mini-PCs was driven by measured performance-per-euro, not preference, and is documented in Lessons Learned.

Technologies

  • K3s on mixed ARM64 and AMD64 nodes
  • Ansible for OS provisioning, K3s install, and ArgoCD bootstrap
  • ArgoCD and Helmfile ApplicationSets for GitOps delivery
  • Longhorn, Garage, and NFS for storage
  • Tailscale, Traefik, and cert-manager for access and TLS

Outcome

The result is a seven-node cluster that can be rebuilt from repositories and whose state is continuously monitored. It has survived hardware replacement, chart deprecations, and a storage migration, which is exactly the kind of operational experience a homelab is for. Everything else in this catalogue runs on it.

Further reading

Alvaro Santos Andres