I also used to run everything on K8s (hybrid Talos / Alpine+k0s on nodes too small for Talos). It is a fun exercise, but with many small nodes eventually I realised I was burning a quarter of my fleet's CPU on just monitoring and routing services.
Eventually I switched to Systemd only, Clan-managed NixOS instances on bare metal.
The killer feature for me is socket activation. It even works with Jellyfin, wrapped in systemd-socket-proxyd. I have Jellyfin listen on a UDS, and Systemd will activate it when my TV connects. The downside is a few more seconds of waiting when I open Jellyfin on the TV. The upside is, after 20 minutes of idling, it is fully powered down.
This allows packing even more services on smaller machines, provided you don't e.g. play Minecraft and watch Jellyfin at the same time.
Eventually I switched to Systemd only, Clan-managed NixOS instances on bare metal.
The killer feature for me is socket activation. It even works with Jellyfin, wrapped in systemd-socket-proxyd. I have Jellyfin listen on a UDS, and Systemd will activate it when my TV connects. The downside is a few more seconds of waiting when I open Jellyfin on the TV. The upside is, after 20 minutes of idling, it is fully powered down.
This allows packing even more services on smaller machines, provided you don't e.g. play Minecraft and watch Jellyfin at the same time.