Raspberry Pi Cluster Recommendation

k3s, k3OS, and k3d
1

I plan to use K3s at home in order to run a HomeAssistant smarthome pod as well as some other relatively tiny pods on it. What would you recommend here?

This is my current resource usage:

NAME                                        CPU(cores)   MEMORY(bytes)   
xy-deployment-8645b48d4f-twxcd              38m          248Mi           
homeassistant-deployment-5c69db6f9c-4nzrw    8m          519Mi           
ab-deployment-7596f7445b-frwlc               4m           23Mi            
asdf-deployment-5c46b55b55-gbdm2             1m           75Mi            
zigbee2mqtt-deployment-6c8cd6c4d8-lr6w4      1m          100Mi
  • A single node K3s “cluster”
  • A three-node cluster with control plane and workload on the same nodes
  • A six-node cluster with three control-plane nodes and three for workloads

Which hardware would you chose? RPi 4 or RPi 5? Would 4 GB RAM be sufficient for the mastern nodes? ( I know SSDs for controlplane nodes are a must to run etcd reliably…)

What are your thoughts?

2

@elronzo does it get busy?
I found the quad core Intel N series are more than ample to run RKE or K3s, I’m using Beelink devices they are NVMe and SSD ready.

3

I would like to keep node costs down to € 100 per node. The Beelink is way over it.

4

@elronzo well on a RPi you still have to add some storage and extra hardware on a RPi for an SSD or NVMe? I don’t see why just one node should be sufficient, I’m using my one for Rancher on RKE2 and rsync backup (to an SSD). It also runs Hyprland on occasions, but most of the time just running/updating and rebooting as I run openSUSE MicroOS on it…

5

The one-node solution was only a thought if costs get out of hand… but maybe a one-node control-plane and three workers… could be a good start.

6

@elronzo well it’s easy enough to add workers if needed and then taint the master node.

You can always look at using Rancher-Desktop (as a single node) deploy your requirements and see how it goes, for example 4 cores and 4GB of RAM to simulate the RPi before committing $$$.