I started running my own infrastructure as a teenager, installing Ubuntu Server 5.10 “Breezy Badger” on a spare computer. Raspberry Pis and older Mac minis followed. The latest iteration spans two locations and supports my home automation, personal tools, and development work.

The big picture

Two locations. One infrastructure.

Location A
4 Proxmox hostsDell OptiPlex Micro1 × 5090 · 3 × 3080160 GB combined RAM
M5 Pro Mac mini64 GB unified memory · 2 TB SSD18-core CPU · 20-core GPU
Synology DS1825+8 hard drives · 16 GB RAM158 TB raw · 111 TB in volumesStorage & containers
Site-to-site VPNUniFi Site Magic
Location B
3 Proxmox hostsDell OptiPlex Micro1 × 5090 · 2 × 3080128 GB combined RAM
M4 Pro Mac mini64 GB unified memory · 2 TB SSD14-core CPU · 20-core GPU
Synology DS1522+5 hard drives · 32 GB RAM160 TB raw · 92 TB in volumesStorage & containers
7 Proxmox hosts2 Mac minis2 NAS464 GB memory200 TB NAS volumes

Memory includes hosts, Mac minis, and NAS. NAS volumes use decimal TB, rounded; 318 TB of raw HDD capacity before RAID and filesystem overhead. SSDs excluded.

UniFi Site Magic links the two networks through a site-to-site VPN. Firewall rules control which traffic can pass between them.

Physical machines

I bought the Dell OptiPlex Micro PCs from Dell Refurbished to experiment with a Proxmox cluster and see how it worked in practice. I upgraded their RAM and storage, fortunately before prices jumped. There are now two 5090s with 64 GB of RAM each and five 3080s with 32 GB each.

The Mac minis handle remote development and local model inference. The two Synology NAS provide storage and also run some containers, including monitoring services.

Virtual machines and containers

Proxmox runs virtual machines (VMs) and LXC Linux containers. Most of my Docker applications run inside VMs, so there are three layers to keep track of: physical host, VM, and application container. Home Assistant has a dedicated VM at each location; Homebridge, Scrypted, Pi-hole, and AdGuard Home use LXC.

Some applications use Docker Swarm, whose scheduler assigns container tasks to nodes. A task can be rescheduled onto another node, so its placement isn’t fixed.

Services

I use the applications below to read feeds, save links and cocktail recipes, control the home, and work with AI tools. Databases, metrics tools, and MCP servers support those applications.

Selected applications, data services, and MCP servers, grouped by purpose.

Homepage brings the service links together. Komodo manages containers across both locations.

Network and access

UniFi gateways and switches connect the machines at each location. Site Magic links the two networks through a site-to-site VPN, with firewall rules controlling traffic between them. The Proxmox hosts and Mac minis use a separate network segment, or VLAN.

Cloudflare Tunnel and Traefik handle application access; AdGuard Home and Pi-hole filter DNS queries. I also use T-Mobile for backup internet access, with download speeds of about 400 Mbps.

Storage and Ceph

Each location has its own Ceph distributed-storage cluster, separate from the NAS. Both use replicated pools configured for three copies of each object: a tradeoff between usable capacity and redundancy.

The Synology DS1825+ has two SHR-1 pools, each tolerating one disk failure; the DS1522+ uses SHR-2 for two-disk tolerance. These are Synology Hybrid RAID configurations. Neither RAID nor replication replaces backups: copies of current data cannot by themselves undo an unwanted change.

For the client side of network storage, I’ve written separately about how I mount SMB shares on macOS.

Backups and monitoring

I use Proxmox Backup Server to back up VMs and LXC containers, with one backup-server VM at each location. I also use Time Machine to back up both Mac minis to the two Synology NAS, one at each location.

Prometheus collects metrics, InfluxDB stores time-series data, and Grafana provides dashboards. Uptime Kuma checks whether services are responding. Knowing I can recover their data takes a successful restore.

Mac minis and inference

Running several Codex and Claude sessions at once was pushing my MacBook Pro’s CPU and GPU to their limits. Using the Mac minis as remote hosts lets me spread that work across multiple machines without leaving my laptop running. I can keep iterating on projects from my iPhone or iPad while I’m out, even when I’ve left the MacBook at home.

Both Mac minis have 64 GB of unified memory and run local model inference. I currently use Ollama on the M4 Pro, but I’m considering a switch to oMLX.

Remote development & inference

One Mac mini at each location.

Development + inference

M5 Pro Mac mini64 GB unified memory · 2 TB SSD18-core CPU · 20-core GPU
CodexClaude
Also runs locally
Local model inferenceRun language models

Development + inference

M4 Pro Mac mini64 GB unified memory · 2 TB SSD14-core CPU · 20-core GPU
CodexClaude
Also runs locally
Local model inferenceRun language models
Both Mac minis run remote development tools and local model inference. Open WebUI runs separately in a VM.

I’m happy with where the setup is now. The next thing I’d like is more headroom for inference: hardware that lets me run larger models.

Explore the infrastructure

The slides below show how the machines, networks, and backups fit together. Choose Read slide for a larger view you can zoom and pan, including on a phone.

Excalidraw / Visual tourMy infrastructure, illustrated

A seven-slide tour of the machines, workloads, services, storage, and backups across both locations.

Open in Excalidraw (new tab)
A spare computer running Ubuntu Server 5.10 led to Raspberry Pis, older Mac minis, and today’s two-location setup.A spare computer running Ubuntu Server 5.10 led to Raspberry Pis, older Mac minis, and today’s two-location setup.
The journey

A spare computer running Ubuntu Server 5.10 led to Raspberry Pis, older Mac minis, and today’s two-location setup.