DiyMediaServer
Featured image of post Homelab Projects Update: Technitium, OpenTofu and 10GbE

Homelab Projects Update: Technitium, OpenTofu and 10GbE

A homelab projects status report: Pi-hole retired for clustered Technitium DNS, OpenTofu drift monitoring, 10GbE on Proxmox, Omada VLANs and NUT UPS shutdown.

Every homelab blog eventually publishes the “sorry I haven’t posted” post. It promises a comeback, lists zero specifics, and is followed by four more months of silence. I’d like to do better than that, mostly because I can tell you I have been very busy with projects. However, a lot of them are half-finished at the same time. It is frustrating for me too.

The blog hasn’t had a new post since early October. That doesn’t mean I stopped working on it. Most of my time went into documenting and automating instead of writing. DNS got replaced. Some guests became code. A node got a new motherboard. The network is just getting started on VLANs with new Omada gear.

So here’s what has been going on. Short status reports, honest about what’s done and what isn’t. The step-by-step instructions come later, starting with Technitium DNS on Proxmox, which is first in the queue.

💭
TL;DR: The blog went quiet but the lab didn’t: Pi-hole is retired in favor of clustered Technitium DNS; OpenTofu now owns my noncritical Proxmox guests; pve-02 escaped its mini-STX case; 10GbE between the nodes is done; Omada VLANs and NUT UPS shutdown are still in progress; and AT&T fiber, with an 8311 SFP module replacing their gateway, is delayed after a botched install. This is a status report, with full write-up posts to follow.

Why the Blog Went Quiet (and the Lab Didn’t)

Writing a guide about a thing requires the thing to stop moving. Nothing in my lab has stopped moving since summer.

Some of these projects started in August, but they all hit the messy middle at once. Finish a stage, find out what it broke, document the fix, discover the next stage needs something I haven’t built. That loop eats evenings and produces no blog posts. I know I should be doing one project at a time, but where is the fun in that?

I built, broke and rebuilt all of it myself on my three-node Proxmox VE cluster. I have used Proxmox Backup Server more times than I would like to admit.

On top of all of this, I’m back in school full-time (and still employed full-time) as someone who is quickly approaching 50. It isn’t as easy as it sounds. I love writing these posts, but they had to take a back seat for a while. I’m hoping that with the holidays around the corner, I can get a few posts published for you all.

DNS: Technitium DNS on Proxmox Replaced My Pi-hole Pair

Status: done.

The Pi-hole pair is gone. So is Nebula Sync, the tool that kept them in sync. In their place are two Technitium DNS LXCs on Proxmox, clustered with Technitium’s native clustering. OpenTofu defines both containers and Ansible configures them. Devices cut over in early September. I finished the migration and deleted the two Pi-hole containers. I now have a grown-up DNS cluster. Yes, it still blocks unwanted ads.

Unfortunate side effect: my Pi-hole sync guide with Nebula Sync now describes a setup I no longer run. It still works for a Pi-hole pair. But you deserve to know I’ve moved on before you take advice from it.

Why a cluster instead of two Pi-holes and a sync job? Fewer moving parts. Technitium’s clustering lets you administer both nodes from one console. Each node still answers queries on its own, so give your clients both addresses.

The part I’m proudest of is the boring part. Once the pair was live, I tested rebuilding it from nothing on isolated throwaway containers. Two rehearsals in early October covered a fresh pair, replacing the second member, and replacing the first after a forced promotion. These were full-blown rebuilds. Across the two rehearsals I found 35 issues, each of which would otherwise have surfaced down the road. It was nice to find them early rather than when I really needed to rebuild because of an equipment failure.

Infrastructure as Code: OpenTofu on Proxmox

Status: done.

Since August, every noncritical guest has lived in code. OpenTofu with the bpg/proxmox provider declares each container and VM, and new VMs are cloned from a Debian 13 golden template. The critical hosts (the router, the backup servers and friends) stay hands-on for now. I’m not brave enough to tackle those yet.

Managing them with code doesn’t stop you clicking around in the Proxmox GUI. And if you do, the code and reality disagree, silently, until the next apply.

Example: I moved four guests off pve-02 by hand, and everything said it was good to go. But it wasn’t good to go. This would have destroyed my Postgres VM, because the bpg/proxmox provider treats moving a VM to another node as a forced replacement (destroy it, then build a new one). Again, a few days later I found a container running 8 cores against a declared 4. Neither caused damage. Both were one careless tofu apply away from it. Make sure you keep OpenTofu up to date with any changes you want to make.

I built a drift monitor to help with this. Four times a day, the controller runs a read-only tofu plan and pages me through ntfy when a managed guest no longer matches its declaration. Does it page me every time I edit a guest in the GUI? Yes. Do I still edit guests in the GUI? Not as much. It is a hard habit to break.

pve-02 Escapes Mini-STX

Status: done. It runs fine.

pve-02 used to be a mini-STX desktop sitting on top of the rack. Cute, and a dead end: two SATA ports, one spare M.2, no PCIe slots, and both SODIMM slots full at the board’s 64 GB maximum.

The fix was a transplant. The socketed i5-13500 moved onto an ASRock B760M board with four DIMM slots, and the node now runs 128 GB of used non-ECC RAM (prices are crazy right now). I ran memtest for 12 full passes across all of it before trusting it. Zero errors.

Then the new storage bit me. Four SATA SSDs became a raidz1 pool, and the pool stalled under load almost as soon as guests landed on it. You’d assume one cause, but no, there were two. I fought with this longer than I should have. I had a bad SATA cable and a bad drive at the same time. I replaced both, threw a 400 GB load test at the pool, and it came through clean.

The Network: Omada Switches, Access Points and VLANs

Status: in progress. Deliberately slow.

For years my lab ran on one flat /24. Not by choice. The old switches were unmanaged, and unmanaged switches can’t do VLANs. Every segmentation plan I drew up was a nice idea, and that was it: a plan and nothing more.

That’s changing. I bought a TP-Link SX3008F 10GbE SFP+ core switch, an SG3218XP-M2 2.5GbE PoE access switch and four EAP775-Wall access points, and the Omada controller is self-hosted in an LXC on pve-02.

Where things stand: only the switches and one access point are adopted. A management VLAN for the Proxmox nodes’ onboard 2.5GbE ports went live. Each node now has a second, isolated path to its SSH and web UI. An admin-workstation VLAN is planned. SSID-to-VLAN mapping, the other three access points, OPNsense trunks and ACLs are still on the list. That’s most of the list. I know.

10GbE Between the Proxmox Nodes

Status: done.

Each node now has an Intel X520, every node’s main bridge runs on it, and iperf3 measured 9.41 Gbps between every pair. Yay!

Getting there wasn’t as easy as I thought it would be. The first card went into pve-01, every link light came on, and a few seconds later all of them went dark. The card was fine. Its PCIe root port enumerated ahead of the others and shoved every other NIC down one bus number, and kernel interface names come from that topology. Every bridge in /etc/network/interfaces now pointed at ports that didn’t exist. pve-01 also hosts my OPNsense router. You can guess how my wife felt about this discovery.

pve-02’s board swap had stranded its networking the same way back in August. Fool me once, shame on the kernel. Fool me twice, shame on me. Every NIC on all three nodes is now pinned to its MAC address with Proxmox’s own interface-pinning tool, and the names survive whatever I plug in next.

AT&T Fiber and an 8311 SFP Module

Status: delayed.

AT&T is installing 1 Gbps fiber, replacing my Xfinity service. I’m happy about the speed. I’m less happy about the gateway AT&T hands you with it, because I already have a router. OPNsense does that job, and it doesn’t need a second box in front of it.

So the AT&T gateway goes in a drawer. The plan is an SFP module flashed with the community 8311 firmware, plugged into an SFP+ port that feeds the OPNsense WAN, so the fiber terminates on hardware I control instead of hardware AT&T controls. One less box on the shelf, and one less device doing NAT before OPNsense ever sees a packet.

It also means retiring a workaround. My OPNsense WAN carries a tunable that shortens the ARP cache lifetime so it re-learns the Xfinity gateway quickly. With no Xfinity gateway, there’s nothing left to re-learn.

None of this has touched a cable yet (see the update below).

Update (November 2026): After digging up my yard, they somehow managed to forget to land fiber at my house and now have to pull a new line. Who knows how long that will take.

NUT UPS Shutdown on a Raspberry Pi

Status: built, not armed.

A Raspberry Pi ran NUT for the rack UPS, and all three Proxmox nodes were NUT secondaries of it. Every piece was in place, and every piece was disarmed on purpose. Then the Pi died (more on that below). A power cut today still ends in hard power loss when the battery runs out.

Why build it and not arm it? Because an untested shutdown system is worse than none. It hands you confidence it hasn’t earned. And I blew up my very old Raspberry Pi 3B (a PoE short) and have a Raspberry Pi 4 on the way.

I meant to use the OBLBDURATION directive to hold each node off on low battery so the Pi’s orderly shutdown gets the first move. It didn’t work. NUT 2.8.1-5 doesn’t know what that directive is, and upsmon logged it as invalid and started anyway. That’s the scary kind of failure: nothing stops and everything looks fine.

Arming is still behind a short list of things I want proven first, ending in an attended battery-pull test. Not a simulated one.

Lightning Round: Smaller Homelab Projects

  • pbs-02 is a second Proxmox Backup Server that pulls from the first. My backups are no longer a single copy.
  • A CUPS print server now shares my USB Brother printer on the LAN. Then I discovered the printer powers itself off after four idle hours, and USB traffic can’t wake it. Jobs sit in the queue until someone presses the power button. Thanks, Brother.
  • Monitor scripts, plural. Certificate expiry, ZFS pool health and SMART are all watched now, with pages through ntfy.

What Broke Along the Way

One lesson kept coming back: redundancy nobody checks isn’t redundancy. Back in August, still on Pi-hole, I found six hosts with no working DNS failover. One was a Pi-hole pointing only at itself. That’s why the Technitium rebuild got tested so hard, and why NUT stays disarmed until a real battery pull proves it.

Frequently Asked Questions

➤ Why would OpenTofu destroy a Proxmox VM I only migrated?
Because the bpg/proxmox provider treats a VM’s node as part of its declaration. Move the VM to another node by hand and the next plan sees a mismatch it resolves by destroying the VM and creating a new one. Nothing warns you until you read the plan, which is why I run a drift monitor.
➤ Does a Technitium cluster mean my clients only need one DNS server?
No. Clustering keeps configuration identical on both nodes, and either node answers queries without the other. It can’t tell a client about a resolver the client was never given. Hand out both addresses through DHCP or static config, or a dead node still takes that client’s DNS with it.
➤ Why is the UPS shutdown built but not armed?
Because an untested shutdown system can fail in ways you won’t see until the power goes out. My first plan relied on a directive my NUT version doesn’t support, and NUT ignored it without stopping. Arming waits for a short proof list ending in an attended battery-pull test.

What’s Coming Next

No dates. I’ve read enough comeback posts to know what a date is worth. What I’ll commit to is the order.

Finished homelab projects get written up first. Technitium DNS on Proxmox, rebuild rehearsals included, leads the queue. OpenTofu and the drift monitor come next, then the 10GbE install. Everything else waits until it’s done.

The lab never stopped.

Sources