apiVersion: apps/v1
kind: Deployment
metadata:
  name: ollama-fallback
  namespace: open-webui
spec:
  replicas: 1
  strategy:
    type: Recreate
  selector:
    matchLabels:
      app: ollama-fallback
  template:
    metadata:
      labels:
        app: ollama-fallback
    spec:
      # Pin the pod strictly to Node 3
      nodeSelector:
        kubernetes.io/hostname: nst-node-3

      containers:
      - name: ollama
        # THE BOOT FIX: Using a frozen tag allows local caching, bypassing the 8-minute pull
        image: ollama/ollama:0.5.4
        imagePullPolicy: IfNotPresent
        # THE FIX: Injects environment variables to pin the model in RAM permanently
        env:
        - name: OLLAMA_KEEP_ALIVE
          value: "-1"
        # THE SPEED FIX: Forces the calculations to fit perfectly onto your hardware threads
        - name: OLLAMA_NUM_THREADS
          value: "4"
        ports:
        - containerPort: 11434
          name: ollama-http
        resources:
          # Matching limits and requests perfectly eliminates CPU scheduling wait loops
          requests:
            memory: "2Gi"
            cpu: "4" # Guarantee 4 physical cores
          limits:
            memory: "4Gi"
            cpu: "4" # Perfectly matches the 4 execution threads above
        volumeMounts:
        - name: local-model-storage
          mountPath: /root/.ollama

      volumes:
      # THE FIX: Bypasses Longhorn replication entirely
      - name: local-model-storage
        hostPath:
          path: /var/lib/ollama-models
          type: DirectoryOrCreate

---
apiVersion: v1
kind: Service
metadata:
  name: ollama-fallback-service
  namespace: open-webui
spec:
  selector:
    app: ollama-fallback
  ports:
  - protocol: TCP
    port: 11434
    targetPort: 11434