feat(ml): physics metrics, Optuna runner, Faz 3 evaluation

infra/k8s/ml/inference.yaml

@@ -0,0 +1,157 @@
+# DeepHorizon inference server — K8s Deployment + Service.
+# Faz 3 Adım 28: gRPC inference + Prometheus metrics.
+#
+# Topology:
+#   - 1 replica (GPU pod, scale=0 ile başla, ihtiyaç olursa artır)
+#   - NodePort 30551 → gRPC (50051)
+#   - ClusterIP 8000 → Prometheus metrics
+#   - GPU: nvidia.com/gpu: 1 (L40S node selector)
+#   - PriorityClass: high-priority (training job'larından önce schedule)
+#
+# Deploy:
+#   kubectl apply -f infra/k8s/ml/inference.yaml
+#
+# Test:
+#   kubectl port-forward svc/inference 50051:50051 -n deephorizon-ml
+#   grpcurl -plaintext -d '{"image": {"data": "...", "mime_type": "image/png"}, "model_id": "pix2pix-v1"}' \
+#     localhost:50051 deephorizon.v1.InferenceService/Enhance
+
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: inference
+  namespace: deephorizon-ml
+  labels:
+    app: inference
+    component: ml-serving
+spec:
+  type: NodePort
+  selector:
+    app: inference
+  ports:
+    - name: grpc
+      port: 50051
+      targetPort: 50051
+      nodePort: 30551
+      protocol: TCP
+    - name: metrics
+      port: 8000
+      targetPort: 8000
+      protocol: TCP
+
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: inference
+  namespace: deephorizon-ml
+  labels:
+    app: inference
+    component: ml-serving
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate  # GPU pod — rolling update sırasında 2 pod aynı anda çalışamaz
+  selector:
+    matchLabels:
+      app: inference
+  template:
+    metadata:
+      labels:
+        app: inference
+        component: ml-serving
+    spec:
+      restartPolicy: Always
+      automountServiceAccountToken: false
+      priorityClassName: high-priority
+      # L40S GPU node selector
+      nodeSelector:
+        nvidia.com/gpu.product: NVIDIA-L40S
+      tolerations:
+        - key: nvidia.com/gpu
+          operator: Exists
+          effect: NoSchedule
+      containers:
+        - name: inference-server
+          image: localhost:32000/deephorizon-inference:v1
+          imagePullPolicy: IfNotPresent
+          args:
+            - "--models-dir"
+            - "/app/exports"
+            - "--port"
+            - "50051"
+            - "--metrics-port"
+            - "8000"
+            - "--max-workers"
+            - "4"
+            - "--log-level"
+            - "INFO"
+          ports:
+            - name: grpc
+              containerPort: 50051
+              protocol: TCP
+            - name: metrics
+              containerPort: 8000
+              protocol: TCP
+          env:
+            - name: PYTHONUNBUFFERED
+              value: "1"
+            - name: PYTHONPATH
+              value: /app
+          resources:
+            requests:
+              cpu: "2"
+              memory: 4Gi
+              nvidia.com/gpu: 1
+            limits:
+              cpu: "4"
+              memory: 8Gi
+              nvidia.com/gpu: 1
+          volumeMounts:
+            - name: models
+              mountPath: /app/exports
+              readOnly: true
+            - name: dshm
+              mountPath: /dev/shm
+          livenessProbe:
+            tcpSocket:
+              port: 50051
+            initialDelaySeconds: 30
+            periodSeconds: 30
+            timeoutSeconds: 5
+            failureThreshold: 3
+          readinessProbe:
+            httpGet:
+              path: /metrics
+              port: 8000
+            initialDelaySeconds: 10
+            periodSeconds: 10
+            timeoutSeconds: 3
+            failureThreshold: 3
+      volumes:
+        - name: models
+          persistentVolumeClaim:
+            claimName: inference-models-pvc
+        - name: dshm
+          emptyDir:
+            medium: Memory
+            sizeLimit: 4Gi
+
+---
+# PVC for ONNX models (MinIO'dan indirilen veya build-time bake edilen)
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: inference-models-pvc
+  namespace: deephorizon-ml
+  labels:
+    app: inference
+    component: ml-serving
+spec:
+  accessModes:
+    - ReadOnlyMany
+  resources:
+    requests:
+      storage: 10Gi
+  storageClassName: microk8s-hostpath

infra/k8s/ml/priority-class.yaml

@@ -0,0 +1,23 @@
+# DeepHorizon PriorityClass definitions.
+# Faz 3 Adım 28: Inference pod'ları training job'larından önce schedule edilir.
+#
+# Deploy:
+#   kubectl apply -f infra/k8s/ml/priority-class.yaml
+
+---
+apiVersion: scheduling.k8s.io/v1
+kind: PriorityClass
+metadata:
+  name: high-priority
+  description: "Inference serving pods — GPU node'larda training'den önce schedule edilir."
+  globalDefault: false
+value: 1000000
+
+---
+apiVersion: scheduling.k8s.io/v1
+kind: PriorityClass
+metadata:
+  name: training-priority
+  description: "Training job'ları — inference yokken GPU node'ları kullanır."
+  globalDefault: false
+value: 100000

infra/k8s/ml/pvcs.yaml

@@ -0,0 +1,14 @@
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: training-outputs-pvc
+  namespace: deephorizon-ml
+  labels:
+    app: ml-training
+spec:
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 50Gi
+  storageClassName: microk8s-hostpath

infra/k8s/ml/training-image-build-job.yaml

@@ -0,0 +1,53 @@
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: training-image-build-20260810-v2
+  namespace: deephorizon-ml
+spec:
+  backoffLimit: 0
+  activeDeadlineSeconds: 1800
+  ttlSecondsAfterFinished: 3600
+  template:
+    spec:
+      restartPolicy: Never
+      automountServiceAccountToken: false
+      initContainers:
+        - name: unpack-context
+          image: localhost:32000/deephorizon-training:2026-08-07b
+          imagePullPolicy: IfNotPresent
+          command: ["/bin/bash", "-lc"]
+          args:
+            - mkdir -p /workspace && tar --touch --no-same-owner --no-same-permissions -xzf /context/context.tar.gz -C /workspace
+          volumeMounts:
+            - name: context-archive
+              mountPath: /context
+              readOnly: true
+            - name: workspace
+              mountPath: /workspace
+      containers:
+        - name: kaniko
+          image: gcr.io/kaniko-project/executor:v1.23.2-debug
+          args:
+            - --dockerfile=/workspace/Dockerfile
+            - --context=dir:///workspace
+            - --destination=registry.container-registry.svc.cluster.local:5000/deephorizon-training:20260810-unet100-v2
+            - --insecure-registry=registry.container-registry.svc.cluster.local:5000
+            - --skip-tls-verify-registry=registry.container-registry.svc.cluster.local:5000
+            - --snapshot-mode=redo
+            - --verbosity=info
+          resources:
+            requests:
+              cpu: "1"
+              memory: 1Gi
+            limits:
+              cpu: "4"
+              memory: 4Gi
+          volumeMounts:
+            - name: workspace
+              mountPath: /workspace
+      volumes:
+        - name: context-archive
+          configMap:
+            name: training-build-context-20260810-v2
+        - name: workspace
+          emptyDir: {}