Ht cloud
ht cloud is the private Kubernetes cluster — the compute substrate behind every service in the ecosystem. Repo: heiervang-technologies/cloud (private). Deploys to ~/ht/cloud are gated on Hei's explicit OK.
Cluster
RKE2 v1.34 with HAMi GPU memory partitioning. Six-node roster:
| Node | Role | GPUs |
|---|---|---|
| Centurion | Control Plane, etcd | 1× RTX 3090 |
| Titan | Control Plane, etcd | 2× RTX 3090 |
| Lithium | Worker | 1× Radeon 780M iGPU |
| Rogue | Worker | 1× RTX 3090 |
| Sentinel | Worker | 1× RTX 3050 OEM |
| Ghost | Removed (DMZ candidate) | — |
How to refresh: kubectl get nodes -o wide from a kubeconfig'd machine. Hardware profile auto-regenerates via cloud/k8s/scripts/update-hardware-profile.
Authoritative sources: cloud/README.md (auto-generated) and cloud/k8s/HARDWARE.md.
Services live on the cluster
Full list is grouped in cloud/README.md. High-level:
- LLM inference — llama.cpp and vLLM instances, model-swapping via omniswap.
- ASR — qwen3-asr on Rogue.
- TTS — trost on Titan; vllm-omni on Rogue.
- Image generation — ComfyUI on Titan;
comfy-openaion Ghost. - Data — PostgreSQL with pgvector, PostgREST, Jena RDF triplestore.
- Observability — monitoring stack per
cloud/README.md.
Constraints
The cluster runs with a deliberately spartan set of platform assumptions:
- No
StorageClass— hostPath only. Every persistent workload declares its own hostPath and node affinity. - No cert-manager.
- LAN / Tailnet-only — no public-exposure path is standardised. Any service that eventually needs one goes through snoop-kube.
These are why every new manifest starts LAN-first with a follow-up commented section for public exposure — see the HT Wiki's eventual cluster manifest for the reference example (wiki.ht.local, hostPath, local CA).
Push discipline
Any commit to ~/ht/cloud can be created freely, but pushes require Hei's explicit OK. Manifest drafts pile up as PRs or held branches; the deploy call is Hei's.
Sources
~/ht/admin/STATE.md— the cluster snapshot table above.~/ht/admin/CLAUDE.md— the cluster overview.cloud/README.md(auto-generated, in-repo) — authoritative live state.cloud/k8s/HARDWARE.md— hardware profile.
See also
- snoop-kube — Chief K8s Engineer; owns manifests and deploys.
- HT Wiki — eventual tenant (
wiki.ht.local).