Four Waves: How a Home Lab Grows Up

A home lab isn’t a static thing. It grows. Not on a plan, not according to a roadmap, but through a series of decisions that each follow the same logic: this worked, so I want to understand what’s next.

I’ve been through four distinct waves over more than a decade. Each one ran alongside a career that was also growing , sometimes the lab was covering ground I wasn’t seeing at work, sometimes it was deepening something I was doing daily, often it was keeping me engaged when I needed more variety than any single job could provide. The curiosity that drove the lab and the depth that came from it weren’t separate things. They compounded each other.

This didn’t happen all at once. It happened a little at a time, across years, because each wave was interesting enough to justify the next one. That’s the only plan there was.

Each wave compounded the previous. The thing I’m running now, Arista switching, MikroTik routing, 40GbE fabric, Kubernetes, Ceph, MaaS provisioning real servers, wasn’t possible without the lessons from a $750 experiment with machines that were already old when I bought them.

Here’s how it happened.

---

Wave One: Make Something Work

It started with curiosity about how cloud worked at the bottom. Not the API layer, not the managed services. The actual mechanism. How does a machine get provisioned? How does hardware go from bare metal to running instance? I’d heard about MaaS, Ubuntu’s Metal as a Service, and I wanted to understand it for real.

So I bought four ancient small form factor machines, a dinosaur HP ProCurve switch, and a Digital Loggers IP power controller strip. About $750 total. The IP PDU was there because those desktop machines didn’t have IPMI, so I needed some way to power cycle them remotely, and Digital Loggers was on the MaaS supported hardware list.

I set it up. I configured MaaS. I commissioned the machines. It worked.

That’s the whole first wave. Make something work. Understand the mechanism. The hardware was slow and the setup was janky and I learned more in that experiment about how bare metal provisioning actually functions than I could have from any amount of documentation.

The IP PDUs will always have a special place in my heart.

---

Wave Two: Make It More Complicated

It worked, so I doubled down. This is how labs grow, not through grand plans but through “that worked, what’s next.”

Wave two was a couple of crappy Supermicro half-depth machines in a rolling rack enclosure in my garage. I lived in the desert. Hot computers are not happy computers. I cut and taped and jury-rigged an AC unit to the enclosure because that was the solution available to me and I wanted the thing to run. It was ridiculous. It worked.

I added a NAS. iSCSI storage, slow but storage. I did my first OpenStack deployment here. My first CloudStack deployment. I built a hilariously useless ESXi setup. I ran enough of these things to start understanding how the pieces actually fit together in practice rather than in architecture diagrams.

Wave two is where the mysteries of how cloud works in the real world started resolving. Not because I read more documentation. Because I broke things, fixed them, broke them differently, and built up the operational understanding that only comes from running real workloads on real hardware that you actually have to care about.

---

Wave Three: Add Rigor

At some point the garage setup hit its limits. I moved to a datacenter eight hours away. Hell of a deal on power and controlled environment. Terrible for the situations where I had to physically go figure something out.

Wave three was about doing it more seriously. 10GbE SFP+. DAC cables, different transceivers, and enough experience with them to see a few fail. Real servers throughout, actual enterprise hardware, not desktop machines pressed into service. More MaaS, this time at a scale that exposed the places where my understanding was shallow.

The Raspberry Pi PXE setup I’d been using for initial installs got retired here. It was cool for what it was but Canonical had already figured out the long-term lifecycle management problem. No need to reinvent it. Knowing when to stop reinventing things and use what already works is a form of rigor too.

This is also where I got serious about what I understood versus what I assumed. Wave one and two had given me a working model of how things fit together. Wave three started exposing the places where that model was incomplete or just wrong. The 10GbE work taught me things about network behavior that 1GbE had quietly hidden. The enterprise hardware taught me things about server management that desktop machines couldn’t have taught me.

Wave three is where the lab stopped being a learning environment and started being infrastructure I had to actually operate. Different discipline. Better lessons.

---

Wave Four: Build a True Datacenter Corollary

Then I decided to just go for it.

40GbE core links. Separate leaf, RJ45, and admin switches to establish strong separation of concerns for access pathways. Switch-level VLAN routing with SVIs for efficiency at the switching fabric layer. Off-lease and refurbished hardware, still, but more recent. GPUs. MaaS. Kubernetes. Ceph. A real attempt at the architecture, not a simulation of it.

The routing moved from UniFi, convenient but limited in what it could teach me, to MikroTik on a Dell R630. Arista switching because I’d never used Arista, the hardware was cheap, it was fast, and it was a chance to learn something I hadn’t learned yet. That’s not a credentials strategy. That’s how you actually build cross-domain depth. You pick the thing you haven’t used and you go figure it out.

Wave four is the platform phase. The lab isn’t a place to experiment anymore. It’s a proof point. It runs real workloads. The DNS record it creates for MaaS is a public DNS record, managed automatically by Ansible. The networking decisions are documented in ADRs. The infrastructure is version-controlled and reproducible. The line between lab and production is intentional, not accidental.

---

What Each Wave Actually Taught

Wave one: the mechanism. How does hardware go from nothing to provisioned? What does the bare metal lifecycle actually look like? That curiosity, answered by doing the thing, is the foundation everything else builds on.

Wave two: the complexity. How do the pieces fit together at scale? What happens when you add storage, add compute, add orchestration? Where does it break and why? The garage rack with the jury-rigged AC unit is embarrassing in retrospect and I’d do it again without hesitation.

Wave three: the rigor. What does it mean to actually operate infrastructure rather than just run it? Where does a shallow understanding become a liability? The eight-hour datacenter put real stakes on the work in a way the garage rack couldn’t.

Wave four: the architecture. What does it mean to build something correctly rather than well enough? What’s the difference between a system that works and a system that’s designed to work? The Arista switches, the MikroTik router, the Ceph cluster. These aren’t just tools. They’re the vocabulary of a real platform.

---

What Wave One Actually Costs Today

The other thing worth saying: you don’t need much to get started.

My MVP today would look something like this, sourced almost entirely from eBay:

• IKEA Lack table : the community calls it a Lack rack. About $19. Fits 10-inch equipment perfectly with some minor modification. Ridiculous and useful.
• Cheap 5-8 port unmanaged switch : under $30. You don’t need managed switching for wave one. You need ports.
• 2-3 half-depth Supermicro servers : the kind designed for pfSense or MikroTik appliance use, with decent CPUs and open RAM slots. Older gear using DDR3 is actually a benefit right now given memory prices. $250 each maximum, free shipping on eBay if you’re patient.
• 1 small form factor desktop : something like a Dell OptiPlex 7050 or similar. $200 on eBay. This is your MaaS node, your utility machine, your first managed host.
• RJ45 Cat5/6 cables : $20. This is not a place to spend money.
• Digital Loggers web power switch : $150 used. Remote power cycling without IPMI. The same brand I used in wave one. They still make them. They still work.

Total: under $1000. Probably under $800 if you’re patient on eBay.

eBay is your friend here. I’ve had good luck sourcing lab hardware from eBay across all four waves. The off-lease enterprise equipment market is large, the prices are reasonable, and the hardware has usually been maintained by someone who cared about uptime. Start there.

Wave one isn’t about the hardware. It’s about commissioning it, breaking it, understanding why it broke, and figuring out what’s next. The $750 I spent a long time ago on ancient SFF machines is about the same as what you’d spend today. The lessons are the same. The curiosity is the same.

---

The Point of All of It

I didn’t build this because I needed to. I built it because building it is how I understand things, and understanding things at this level is what makes the advice I give worth taking.

The organizations I work with are running real infrastructure under real conditions with real consequences. The credibility to advise them comes from having run real infrastructure under real conditions with real consequences. Not enterprise infrastructure with a team and a budget. A platform built in waves, with $750 of curiosity money, a jury-rigged AC unit in a garage, and a willingness to keep going when each wave exposed how much there still was to learn.

Wave four isn’t the end. It’s just the current state of a system that keeps growing because there’s always something next.

The MaaS node is commissioning right now. The DNS record just got created automatically. The cluster bootstrap is next.

It worked. So I’m doubling down.