Architecture

DC Suite is a set of services that run on Kubernetes and drive a bare-metal GPU fleet through a fleet manager.

Components

flowchart TB
    subgraph K8s[Kubernetes cluster]
        UI[Web console]
        API[Control-plane API server]
        RUN[Provisioning runner]
        WF[Workflow engine]
        DB[(PostgreSQL)]
    end
    subgraph Shared[Shared services]
        SSO[Identity provider / SSO]
        OBS[Metrics + logs backends]
    end
    FM[Fleet manager]
    POOL[(Bare-metal GPU pool)]

    UI --> API
    API --> DB
    API --> WF
    WF --> RUN
    API --> FM
    RUN --> FM
    FM --> POOL
    API --> SSO
    API --> OBS

Component	Role
Web console	The React single-page app users and operators interact with.
API server	The control plane: REST API, authz, cluster lifecycle, catalog.
Workflow engine	Runs durable, multi-step operations (create, resize, delete) reliably.
Provisioning runner	Executes per-run infrastructure jobs that provision nodes.
PostgreSQL	System of record for clusters, nodes, images, templates, tenancy, cost.
Fleet manager	Owns the physical machines: discovery, power, validation, allocation.
Identity provider	SSO for all sign-in (no local passwords).
Metrics + logs backends	Time-series metrics and log aggregation for observability.

Request flow: creating a cluster

sequenceDiagram
    participant User
    participant API
    participant WF as Workflow engine
    participant FM as Fleet manager
    participant Runner
    User->>API: POST /v1/clusters
    API->>WF: Start create workflow
    WF->>FM: Reserve N nodes (profile)
    FM-->>WF: Reserved
    WF->>FM: Power on + boot golden image
    WF->>Runner: Provision + configure nodes
    Runner-->>WF: Nodes ready
    WF->>API: Cluster READY
    API-->>User: Operation complete

The API returns an operation id immediately; the workflow engine drives the rest to completion (or a clean failure) even across restarts.

Fleet integration

The fleet manager is the authority on physical machines. DC Suite:

1. Reserves machines of a profile from the pool.
2. Asks the fleet manager to power on and validate them.
3. Provisions them from a golden image and forms the cluster.
4. On delete, releases them — the fleet manager sanitizes and re-validates each machine before returning it to AVAILABLE.

A periodic fleet sync reconciles DC Suite's node table with the fleet manager's inventory: machines the fleet reports available become selectable; machines it reports assigned/sanitizing/unhealthy are quarantined so they're never handed out by mistake.

Data model (high level)

• Organizations → tenants → clusters → nodes.
• Images and templates (stacks) are catalog objects.
• Usage records feed cost rollups.
• Bare-metal machines are the operator's registration of physical hardware, linked to fleet-manager machine IDs.

Security posture

• SSO-only sign-in; a separate operator door gates admin access.
• Scoped RBAC: roles grant permissions at installation / org / tenant scope.
• Per-tenant isolation across clusters, logs, and cost.
• Short-lived cluster access: SSH keys are delivered to nodes and their application is confirmed; credentials aren't long-lived secrets baked into images.

---

Next: Installation.