Skip to main content

Kubernetes Cluster Orchestration

Business Value: Go from bare metal to a fully functional, GPU-enabled Kubernetes cluster — with networking, monitoring, and GPU operators pre-configured — through a simple portal form.

How It Works​

Users select their desired configuration: Kubernetes version, CNI plugin, node pools (control plane + GPU workers), host flavors, labels, and taints. The platform handles everything else:

  • Installs container runtime, Kubernetes binaries (kubelet, kubeadm, kubectl), and bootstraps the control plane with etcd in HA quorum (3 or 5 nodes)
  • Deploys the chosen CNI plugin for pod networking — supporting both BGP-based and eBPF-powered options
  • Joins GPU worker nodes, applies labels and taints, and deploys GPU runtime operators: GPU drivers, device plugin, telemetry exporter, and GPU feature discovery
  • Runs conformance checks: API server reachable, all nodes Ready, CNI healthy, DNS resolving, GPUs visible

Kubernetes Cluster Architecture

Technical Highlights​

  • All container images pulled from the platform's internal registry — enabling air-gapped, version-controlled deployments
  • Static pods for control plane components (apiserver, etcd, controller-manager, scheduler) — self-bootstrapping without external dependencies
  • GPU runtime operators deploy as DaemonSets: device plugin registers GPU resources per node, telemetry exporter provides per-GPU metrics, feature discovery labels nodes with GPU properties
  • Full lifecycle operations: upscale, downscale (cordon → drain → remove), rolling upgrades (zero-downtime), and clean deletion with host return to pool
  • Multi-vendor GPU support: compatible with GPU operators across major GPU vendors and extensible to additional GPU plugins

Powered by CKP (Coredge Kubernetes Platform)​

Dflare AI uses CKP (Coredge Kubernetes Platform), a custom Kubernetes distribution built and maintained by Coredge.io. CKP takes upstream Kubernetes source code and produces enterprise-grade, digitally signed binaries and container images, providing a fully supported, supply-chain-verified Kubernetes platform.

Two-Layer Architecture​

  • CKP Distribution Layer — Custom-built Kubernetes binaries (kubeadm, kubelet, kubectl) tagged and signed by Coredge, with Coredge-hosted core component container images.

  • CKP Management Layer — Cluster lifecycle management through CAPI (Cluster API) with Kamaji as the hosted control plane provider.

Supported Versions (CNCF Certified)​

Kubernetes VersionCNI Version
1.28.33.27.2
1.29.0, 1.30.63.28.2
1.31.23.30.5
1.32.113.30.5
1.33.73.30.5
1.34.33.30.5
1.35.13.30.5
ComponentSupported
OSUbuntu 20.04, Ubuntu 22.04
ArchitectureAMD64, ARM64

Key Capabilities​

  • PGP-signed package distribution via BYOH (Bring Your Own Host) bundles
  • Automated cluster provisioning through portal UI and API
  • Ceph-backed persistent storage (ckp-block storage class)
  • Velero backup and disaster recovery
  • TLS certificate management with 10-year validity