notes
Compute: five nodes, containers first
The problem
A single big server is simple until I need to reboot it. I wanted to patch a host, move its workloads elsewhere, and have nothing user-facing notice the difference.
The shape
5 clustered nodes → shared config + quorum → system containers (dense, default) → full VMs (only where a distinct kernel is required)
The hard part
Containers share the host kernel, which is what makes them dense enough to run this many on modest hardware — and also what makes the choice consequential. Anything needing its own kernel, a different operating system, or genuine isolation has to be a full VM, and deciding wrong means migrating a running service later. The default is a container; the exceptions are deliberate and few.
The second lesson was about counting. For a long time the dashboard reported three nodes because the host list was hardcoded and two nodes had joined since. The infrastructure was correct and the page describing it was wrong — which is the more embarrassing failure, and the reason the numbers in the header of this site are now derived rather than typed.