Docker on Packets & Regrets

The Website Was New. The CSS Was Four Months Old.

Fri, 31 Jul 2026 22:00:00 +0200

The rebuilt site passed its Docker build. The origin container was healthy. Every required route returned HTTP 200. Desktop and narrow mobile screenshots from the production-equivalent preview looked correct.

Then I opened the public site on a phone.

It looked awful.

The background belonged to the previous dark theme. Product cards had lost their borders and structure. Labels collided. Sections that were carefully spaced in preview ran into each other in production.

What Actually Runs in the Home Lab

Sun, 19 Jul 2026 20:20:00 +0200

At the publication snapshot, the two Docker hosts were running 76 containers: 61 on the general application host and 15 on the specialist host.

That is not 76 independent products. Sidecars, exporters, databases, browser workers, VPN helpers, and document processors inflate the container count. This inventory groups them by the job they perform instead of pretending every YAML service deserves its own logo wall.

It also omits exact addresses, internal names, camera topology, credentials, firewall details, and management paths. A service catalogue should explain the system, not pre-write somebody else’s reconnaissance report.

Why the Lab Is Boring on Purpose

Sun, 19 Jul 2026 20:10:00 +0200

The current lab is mostly Docker Compose, HAProxy, PostgreSQL, ZFS, and shell scripts.

That sentence would have embarrassed an earlier version of me. The previous version ran Talos Kubernetes with Flux because Kubernetes was interesting, GitOps was fashionable, and apparently a family photo library needed a control plane.

The cluster worked. It also turned every small operational question into a distributed-systems question. Eventually the useful lesson became obvious: a home lab should optimize for the operator’s time, not for the number of CNCF logos in its diagram.

Prometheus on NFS: Healthy Until It Wasn't

Sun, 19 Jul 2026 19:30:00 +0200

The monitoring page was green. The container was healthy. Queries worked.

The storage design was still wrong.

I had put the Prometheus time-series database on an NFS share. It was roomy, already mounted, and moved roughly 30 GiB away from a Docker VM whose local disk was tight. This seemed like a reasonable home-lab compromise right up to the moment the NAS stopped answering.

Then Prometheus became the incident it was supposed to help explain.