Skip to main content

Source: oceanid/docs/architecture/overview.md | ✏️ Edit on GitHub

Oceanid Cluster Architecture

Network Topology​

Calypso Access Path (GPU)​

Component ownership:

  • Pulumi HostCloudflared ensures systemd unit + config on Calypso.
  • Pulumi HostDockerService ensures tritonserver.service with GPU flags.
  • DNS gpu.<base> CNAME is created when NodeTunnels are disabled (host connector path) or by NodeTunnels component otherwise.

Traffic Flow​

Security Architecture​

Component Overview​

Core Infrastructure​

  • K3s: Lightweight Kubernetes distribution
  • Cloudflare Tunnel: Secure ingress without exposed IPs
  • Gateway API: Modern ingress management (2025 standard)
  • cert-manager: Automatic TLS certificate management

Node Configuration​

NodeRoleIPProviderResourcesPurpose
TethysControl Plane157.173.210.123Hostinger2 vCPU, 2GB RAMMaster node
StyxWorker191.101.1.3Hostinger2 vCPU, 2GB RAMApplication workloads
MeliaeWorker140.238.138.35Oracle Cloud1 OCPU, 1GB RAMLight workloads
CalypsoGPU Worker192.168.2.68LocalRTX 4090, 32GB RAMML/AI workloads

Security Features​

  • Zero-Trust Architecture: No implicit trust
  • Network Segmentation: Namespace isolation
  • Automatic Rotation: 90-day SSH, 90-day TLS, 365-day K3s
  • External Monitoring: Sentry (no cluster overhead)
  • Firewall Layers: Cloudflare WAF + UFW + NetworkPolicies

Infrastructure as Code​

  • Pulumi: TypeScript-based IaC
  • ESC: Environment, Secrets, Configuration management
  • GitOps Ready: Prepared for ArgoCD/Flux integration
  • Automated Provisioning: Node joining automation

Scalability Design​

Deployment Pipeline​

Resource Utilization​

Current Usage (Baseline)​

  • Control Plane: ~500MB RAM, 0.5 CPU
  • Worker Nodes: ~300MB RAM each, 0.2 CPU
  • Total Cluster: ~1.4GB RAM, 1.1 CPU

With Monitoring​

  • Sentry Relay: 50MB RAM, 0.05 CPU
  • Error Collector: 32MB RAM, 0.01 CPU
  • Health Reporter: 32MB RAM, 0.01 CPU
  • Total Monitoring: ~114MB RAM, 0.07 CPU

Reserved for Applications​

  • Available RAM: ~6GB across cluster
  • Available CPU: ~4 cores
  • GPU: RTX 4090 (dedicated for ML)

High Availability Considerations​

Current Limitations​

  • Single control plane (Tethys)
  • No etcd backup automation
  • Manual disaster recovery

Future Improvements​

  1. Add HA control plane nodes
  2. Implement Velero for backups
  3. Multi-region deployment
  4. Automated failover

Network Segmentation​