feat: Weight Controller skeleton that watches ExternalModels

[yehuditkerido]

Create the complete weight-controller infrastructure: Go module, controller skeleton, deployment manifests. The controller watches ExternalModel resources and logs reconciliation events.

What's Included:

• Go module structure (weight-controller/)
• Import ODH Payload Processing types (ExternalModel, ExternalProvider)
• Controller-runtime based controller that watches ExternalModel
• Logs when ExternalModel is created/updated/deleted
• RBAC manifests (ServiceAccount, ClusterRole, ClusterRoleBinding)
• Deployment manifests (Deployment, Service)
• Dockerfile (multi-stage, distroless base)
• Makefile with build, test, lint, docker-build targets

Directory Structure:

# Go module (follows maas-controller pattern)
weight-controller/
├── go.mod
├── go.sum
├── Makefile
├── Dockerfile
├── cmd/
│   └── main.go
└── pkg/
    ├── controller/
    │   └── weight/
    │       ├── controller.go      # Main reconciliation loop
    │       └── calculator.go      # Weight calculation algorithm
    │
    ├── metrics/
    │   ├── types.go               # ClusterMetrics struct
    │   ├── source.go              # ClusterMetricsSource interface
    │   │
    │   ├── vllm/                  # VRAM utilization (from vLLM)
    │   │   └── scraper.go         # vLLM Prometheus scraper (:8000)
    │   │
    │   └── epp/                   # Latency (from llm-d EPP)
    │       └── scraper.go         # EPP Prometheus scraper (:9090)
    │
    ├── health/
    │   └── tracker.go             # Cluster health state tracking
    │
    └── config/
        └── config.go              # ConfigMap-based configuration

# Deployment manifests (centralized, follows repo pattern)
deployment/base/weight-controller/
├── default/
│   └── kustomization.yaml      # ties everything together
├── rbac/
│   ├── kustomization.yaml
│   ├── serviceaccount.yaml
│   ├── clusterrole.yaml
│   └── clusterrolebinding.yaml
├── manager/
│   ├── kustomization.yaml
│   └── deployment.yaml
└── monitoring/
    ├── kustomization.yaml
    └── service.yaml

# Sample CRs
docs/samples/weight-controller/
└── externalmodel-multi-cluster.yaml

Architecture:

        vllm/scraper.go              epp/scraper.go
        (VRAM utilization)           (Latency)
              │                           │
              └───────────┬───────────────┘
                          │
                          ▼
                   ClusterMetrics
                          │
                          ▼
               controller/weight/
               calculator.go
                          │
                          ▼
               Patch ExternalModel

Acceptance Criteria:

• make -C weight-controller build succeeds
• make -C weight-controller docker-build creates image < 100MB
• kubectl apply -k deployment/base/weight-controller/default/ deploys successfully
• Controller logs "Reconciling ExternalModel: " when ExternalModel is created
• Controller reconnects after API server restart

Demo:

# Deploy controller
kubectl apply -k deployment/base/weight-controller/default/

# Create an ExternalModel with multiple providers
kubectl apply -f docs/samples/weight-controller/externalmodel-multi-cluster.yaml

# See controller logs
kubectl logs -l app=weight-controller -n maas-system
# Output: "Reconciling ExternalModel: my-model"

[yehuditkerido]

/assign