Kubernetes on Hetzner: cutting my infra bill by 75%

Original Article ↗ Hacker News Discussion ↗

Cost vs Operational Trade‑offs

  • Multiple commenters report infra bills on Hetzner being ~20–25% of AWS for equivalent capacity, especially when bandwidth or storage dominate costs.
  • Others emphasize TCO: self‑managing Kubernetes + storage (Ceph, etc.) can become a full‑time DevOps job and may wipe out savings, especially after outages.
  • Debate over when it becomes cheaper: some argue once cloud spend hits roughly mid–five figures per month and you have at least a couple of strong infra engineers, Hetzner/bare metal wins; others say even short outages can negate savings.

Storage and Databases on Hetzner

  • Strong consensus that Hetzner cloud volumes are too slow for serious production databases; high IOWAIT and low IOPS are common.
  • Suggested mitigations:
    • Use bare‑metal nodes with local NVMe (often RAID10).
    • Run DBs outside K8s on metal, or use K8s with local NVMe and node pinning.
  • Ceph (rook‑ceph) is seen as powerful but complex and often poor value at small scale; some prefer simpler NFS or block‑replication setups.

Cluster Provisioning & Tooling

  • Popular tooling mentioned: terraform‑hcloud‑kube‑hetzner, Cluster‑API + Hetzner provider, Talos + Omni, k3s, and various operators (DB, MinIO, load balancer).
  • Some vendors offer “managed Kubernetes on Hetzner” layers to provide self‑healing and one‑click upgrades while still benefiting from low prices.

Hybrid / Multi‑Environment Clusters & Networking

  • Several people explore clusters spanning on‑prem + cloud or multiple providers.
  • Techniques: WireGuard overlays, Tailscale operator, Cilium, Nebula, Netmaker, BGP on Hetzner vSwitch, etc.
  • Skeptics warn that extra hops, asymmetric routing, and “internet‑quality” links can wreck performance during peak load; others counter that with good design (edge caching, peering DCs) it can work.

Reliability, Support, and Abuse Handling

  • Experiences with Hetzner support range from “outstanding, very direct and technical” to “they null‑routed us on launch day and took days to fix.”
  • Reports of vSwitch resets, false‑positive abuse triggers, and fair‑use limits on “unlimited” 1 Gbit traffic.
  • Some see this as acceptable trade‑off for price; others prefer AWS/major clouds for more predictable support and fewer surprise interventions.

Kubernetes Complexity & Alternatives

  • Multiple voices question using Kubernetes at small scale, calling it overkill compared to simpler schedulers (e.g., Nomad) or even basic VMs/compose.
  • Counter‑arguments: even single‑server k3s can pay off where cloud is expensive; K8s APIs (Ingress, Services, PVCs, CRDs) and ecosystem (operators, Helm) solve many hard problems cleanly.
  • General agreement: K8s adds significant complexity; managed control planes or expert help are often worthwhile.

Hetzner vs Other Providers and Environment

  • Hetzner is consistently seen as far cheaper than DigitalOcean, OVH, and orders of magnitude cheaper than AWS egress.
  • Some worry about IP reputation (blacklisting, email deliverability) typical of budget providers.
  • Sustainability briefly discussed: EU Hetzner DCs are said to use certified renewable energy; US locations are unclear. Some argue data‑center emissions are non‑trivial and should be considered; others see transport and other sectors as much higher‑leverage targets.