← Holocron Logs

Writing ADRs for a Homelab: Ten Decisions, Zero Amnesia

Architecture Decision Records sound like enterprise ceremony until you forget why you made a choice six months ago. The Alliance Fleet now runs on ten ADRs, and they have already changed how the lab evolves.

Why this matters beyond the homelab: ADRs are standard practice on cloud and platform engineering teams. Practicing the discipline of documenting context, decision, and consequences maps directly to how infrastructure teams operate at scale.


The Problem ADRs Solve

Every infrastructure decision has two lifespans. The first is the moment you make it, when the context is vivid and the tradeoffs are obvious. The second is six months later, when you are staring at a config and asking why past-you did something that looks wrong.

Without a record, you get one of two bad outcomes. Either you re-litigate the decision from scratch and waste an evening rediscovering the constraint that forced it, or you “fix” it and rediscover the constraint the hard way, in production.

An Architecture Decision Record is a short document that captures three things: the context at the time, the decision made, and the consequences accepted. That is the entire format. The Alliance Fleet Codex now holds ten of them on a dedicated BookStack shelf.


What a Homelab ADR Actually Looks Like

Two examples from the fleet show the range.

ADR-008: VM placement policy. Which workloads belong on which Proxmox node, and why. Node-B carries the data and identity tier because it has ECC memory and ZFS. When I later found two VMs living on the wrong node, the ADR turned a judgment call into a compliance check: the document said where they belonged, so the July migration of Home-One and Coruscant to Node-B was not a debate, it was a work item.

ADR-010: Local model selection for agent workloads. The original decision selected qwen3:32b as the default model for Jocasta, the fleet’s AI agent. Reality intervened: 32b exceeded the 20GB VRAM on the RTX 4000 SFF Ada. Instead of silently swapping models, I wrote an addendum documenting the substitution to qwen3:14b, the measurement that forced it, and what capability was traded away.

That addendum is the part worth internalizing. ADRs are not carved in stone. When a decision gets revised, the record grows instead of being rewritten, and the reasoning chain stays intact.


The Discipline, Not the Template

The template is trivial. Context, decision, consequences, status. What matters is the habit:


Key Takeaways

← Back to Holocron Logs