Kubernetes Distributions

Hosted control plane (cloud)

★ EKS (AWS) — most mature; control plane $73/mo + nodes; deeply integrated with IAM.
★ GKE (GCP) — best DX of the cloud k8s; Autopilot mode is hands-off; great Cilium support.
AKS (Azure) — solid; AD integration; free control plane on certain tiers.
DigitalOcean Kubernetes (DOKS) — free control plane, simple, cheap nodes.
Linode Kubernetes Engine — same niche as DOKS; free control plane.
Vultr Kubernetes Engine — cheap.
Civo — k3s under the hood; free control plane, fast spin-up.
Scaleway Kapsule — EU; competitive.
OVH Managed Kubernetes — EU; free control plane.
Hetzner Cloud + k3s — DIY but cheap; many run this combo.

★ kubeadm — official "build your own cluster" tool. Most learning, most control, more ops.
★ k3s (Rancher / SUSE) — single-binary "Kubernetes lite" that's actually full upstream Kubernetes. Default for edge / small clusters / homelab.
k0s (Mirantis) — competitor to k3s; single binary, simpler upgrade story.
Talos Linux (Sidero) — immutable, API-driven, no SSH; declarative cluster spec. The modern self-host pick for serious workloads.
Rancher RKE2 — SUSE's Kubernetes distro; production-hardened.
Cluster API (CAPI) — Kubernetes as the API for managing other Kubernetes clusters.
kops — older AWS-focused cluster bootstrapper; less common.
Charmed Kubernetes (Canonical) — for Ubuntu shops.
OpenShift (Red Hat) — k8s plus a lot of extras; commercial.

k3s — also great here; tiny footprint.
MicroK8s (Canonical) — single-package Ubuntu k8s; addons (DNS, ingress, storage) one toggle each.
k3d — k3s in Docker; very fast for local; not for prod.

See Local Kubernetes — Kind, k3d, Minikube, Rancher Desktop.

See Container Host OS — Talos, Bottlerocket, Flatcar, Fedora CoreOS.

Cluster API (CAPI) + a provider (CAPA / CAPG / CAPV) — declarative cluster lifecycle.
Talos talosctl upgrade — single command for OS + k8s.
k3s Install_K3s_VERSION=… curl … — re-run installer to upgrade.
Kured — coordinated reboots after kernel updates.
kubeadm upgrade — manual but well-documented.

Don't run your own etcd unless you must. Hosted distros and Talos handle it.
Pin k8s minor versions in IaC. Auto-upgrades to a new minor break things.
At least 3 control-plane nodes for HA self-host (or single-node-with-backups if you accept the risk).
Separate control plane and workload nodes in larger clusters.
Have a recreation plan. Cluster gone → 30-min rebuild via Cluster API or Talos apply.