CERBERUS 2.0
Lean upstream Kubernetes on Debian that begins as one useful hybrid node, adds workers quickly, and expands the control plane only when the cluster crosses explicit topology tiers.
Problem
Small Kubernetes deployments are often forced to choose between an overbuilt high-availability control plane and a disposable single-node lab. CERBERUS 2.0 keeps the first node useful by itself while preserving a clean path into a resilient multi-control-plane cluster.
Constraints
- The bootstrap node must remain both control plane and schedulable worker.
- Worker capacity should arrive before control-plane expansion finishes.
- Stacked-etcd membership should remain odd and topology changes should happen only at profile thresholds.
- Scale-down, demotion, and failed retirement must remain observable and retryable.
Architecture
- 01
kubeadm, containerd, Cilium, CoreDNS, and kube-vip provide a deliberately upstream-leaning Debian substrate.
- 02
A stable API VIP exists from the first boot so later control-plane growth does not require rebuilding cluster identity.
- 03
Data-driven profiles map enabled-node counts to desired hybrid control-plane counts, with sticky membership between tier transitions.
- 04
Scale-out joins workers first, then serially promotes eligible nodes; scale-down drains and resets retiring nodes before recalculating the tier.
Failures / lessons
- High availability is a topology decision, not a box to check before the workload exists.
- Odd etcd membership avoids paying replication cost for no additional failure tolerance.
- Scale-down deserves explicit state because demotion exposes coupling that scale-out can hide.