Overview
Atlas is my homelab: something I’m building as a hobby and to learn DevOps properly, by running real infrastructure instead of following tutorials.
Work in progress: this page tracks where it’s up to.
Hardware
atlas1: a Dell OptiPlex 7090 Micro with a 10th-gen Intel Core i7, 32 GB of RAM and about 930 GB of storage.
Current state
No services yet. Right now atlas1 runs a single OPNsense VM, the firewall and router that everything else will sit behind.
Design principles
The plan behind Atlas has been through three architecture audits, and a few rules shape everything in it:
- Everything as code. Infrastructure is defined in OpenTofu and Ansible and deployed from git, so any machine can be rebuilt rather than repaired.
- Criticality tiers. Production, homelab services, learning environments and experiments are kept apart, and a higher tier never depends on a lower one. An experiment crashing can’t take down anything people rely on.
- The internet gateway doesn’t depend on the hypervisor. The router eventually moves to its own box, so rebooting a server never takes the network down.
- Backups only count once restored. Every backup target has a written recovery procedure and a drill.
Roadmap
The plan grows in stages, each one tied to what it should teach:
- Everything as code (now). Import the existing setup into OpenTofu and Ansible; a three-node k3s cluster with Cilium for learning etcd quorum and eBPF networking; monitoring, alerting and off-site backups.
- Dedicated router. OPNsense moves off the VM onto a fanless mini PC with a managed switch, carrying VLANs on real ports.
- Memory and backups. More RAM, Proxmox Backup Server and single sign-on with Authentik.
- Platform engineering. Two more nodes for a three-node Proxmox cluster; Kubernetes rebuilt with kubeadm and deployed through Argo CD (GitOps); distributed storage with Longhorn; short-lived secrets with OpenBao; image scanning and signing in CI.
- Lab node, GPU and robotics. A GPU worker for ML pipelines, ROS 2 simulation as a CI stage, and a small Raspberry Pi edge fleet.
- Network and source of truth. Full VLAN segmentation, self-hosted NetBird for remote access, and NetBox as the source of truth for addresses and devices.
- NAS and recovery. A ZFS NAS, immutable off-site backups, and a written, drilled recovery runbook.
- Smart home. Home Assistant on a dedicated production node with automatic failover.
- Security operations. Cameras on an isolated network, plus Wazuh, Falco, canary decoys and supply-chain security in CI.
- AI server. Locally hosted LLM agents that can only act through a policy gateway with approvals and an audit log.
- Fabrication and robotics. 3D printers, a mobile robot and a robot arm, tied into the lab’s network and simulation.
- FPGA lab. A hardware order-book pipeline on a DE10-Nano, run on paper data.
Stack (current and planned)
Proxmox · OPNsense · OpenTofu · Ansible · Kubernetes (k3s → kubeadm) · Cilium · Argo CD · Prometheus & Grafana