I'm maintaining a nomad cluster already a few years now at home, based on some thin clients and a few raspberry pi's.

The workload is growing from uses cases of the plane-spotting services towards a pi-hole setup, vaultwarden, homeassistant and many more use cases.

One of the issues I encountered was based on the default scheduling algorithm. Raspberry pi's are not known as the most efficient solution to run a huge workload. Default nomad will schedule new containers on one compute node until the resource limits of that node are consumed and only then will start consuming another node. This …