Resource Planning

Resource Planning

My five node cluster currently consists of:

1. Three control plane nodes (cp1, cp2 and cp3: 16GB ram Raspberry Pi 5s): all running Raspberry Pi OS (Debian 13) in headless mode so are lightweight and Pi optimised. Etcd requires the most RAM (3GB for small clusters like mine with <100 pods), while other components should typically stay under 1GB each (with Prometheus at roughly 3gb as an exception) with my multi-node control plane set up which distributes load, so 16GB handles 20-50 pods across nodes without OOM issues.
2. Two worker nodes (wn1, wn2), one runs Debian and the other Ubuntu.

I have mixed 8GB and 36GB worker nodes which work fine for my Kubernetes homelab, as schedulers automatically balance pods based on allocatable resources (node taints/labels also help). 8GB should handle 20-40 pods (e.g., Prometheus targets, Home Assistant), while 36GB can take heavier workloads like databases or future AI experiments without OOM issues.​

Having two control plane nodes promotes High Availability (H.A) for the things I will run on my cluster, which would be frustrating to have offline:

• HomeAssistant
• AdGuard (home dns server)

To achieve H.A in a real world production environment using a hybrid cloud/on premise or cloud/cloud (eg aws/gcp) setup would be obviously more sensible than a cluster of raspberry pis. Also, my control plane nodes share the same ip via kube-virtual-ip which is owned by one of my control plane nodes. In a production environment this would be its own dedicated machine or better still a load balancer instead of kube-vip. In order to achieve High Availability (H.A) the minimum number of nodes in an etcd cluster is 3. This is to achieve quorum which cannot be achieved with less than 3 nodes and to obtain a failure tolerance of 1. You can read more here: https://etcd.io/docs/v3.3/faq/#why-an-odd-number-of-cluster-members

I will create two namespaced environments, dev and prod. No pod in either environment runs on a control plane node.