Build note
Docker became the default home for services
The easy part of self-hosting is starting a container. The hard part is remembering which of the folders are running, which ports are public on the LAN, and which databases should never be reachable from outside Docker.
The Compose files live under ~/aqrtr/docker-files/, one service or stack per
directory. The inventory captured forty service directories. That is not forty
containers running. The P920 has a much smaller set of active containers, and
the set changes as I turn projects on and off.
One directory per stack
Compose gives each service a small home: a YAML file, its startup settings and the volumes it needs. A few examples from the inventory are:
| Service | Purpose | LAN port in the inventory |
|---|---|---|
| Gitea | Self-hosted Git | 3087 |
| SearXNG | Search | 17327 |
| Glance | Home dashboard | 14526 |
| Portainer | Container management | 39443 |
| cAdvisor | Container metrics | 50084 |
The host also runs services outside Docker. Prometheus, Grafana and the NVIDIA exporters are system services, not Compose projects. Drawing that line matters when Docker is unhealthy: a host-level Prometheus can still tell me whether the machine is alive even if the Docker API is not answering.
Ports are part of the design
The published ports are high and explicit. Redis and databases that only serve
other containers stay on an internal network or bind to loopback. For example,
the local Redis service binds to 127.0.0.1:6379. A port map is not just a
convenience; it is the short list of doors this machine opens to the LAN.
The reverse proxy can provide a friendlier address, but the raw port map still matters for diagnosis. If the dashboard is down, I want to know whether the container stopped, the host port changed, or the proxy is the part that failed.
What goes wrong when the daemon restarts
One Docker restart looked stuck because docker.service remained in
activating (start). It had no useful start timeout. The daemon was actually
re-attaching its Compose networks and cleaning up stale containerd shims; it was
slow and unhealthy, but the evidence did not support calling it a deadlock.
The lesson was to inspect the systemd state and Docker journal before issuing a
second restart. Repeated docker start and docker rm -f commands had their own
clients hang while the daemon was still reconciling its state. A command that
does not return is not proof that the daemon made no progress.
The rule that would have prevented the next week
Keep the Compose folder map and the port map together, and distinguish “defined” from “running.” Keep data stores internal unless a client needs a host port. When Docker is slow, inspect its state once before stacking more commands on top of it.
How this was made
- Written by: Rupesh, from his own Compose folders and container inventory, with AI drafting and editing.
- Ports and service names: from the rtr-p920 inventory snapshot. The running set changes over time.
- AI-assisted sections: structure and copy editing.