ai-inference-service-mesh

AI 推論サービスメッシュ

Istio／Linkerd メッシュの制御を適用し、推論マイクロサービス間の東西方向の AI トラフィックを保護して最適化します。

AI にメッシュを導入する理由

• ゲートウェイ、リトリーバー、リランカー、モデルサービス間で mTLS を強制します
• アプリケーションコードを変更せずに、きめ細かなトラフィックポリシーを適用します
• モデル提供バックエンドで段階的なデリバリーを実行します
• 検索＋生成チェーンにおけるレイテンシーのホップを可視化します
• モデルのバージョン、テナント、優先度層に基づいて推論リクエストをルーティングします
• 高価な GPU ベースのサービスを連鎖的な障害から保護します

前提条件

# Install Istio with production profile
istioctl install --set profile=default \
  --set meshConfig.accessLogFile=/dev/stdout \
  --set meshConfig.defaultConfig.holdApplicationUntilProxyStarts=true

# Label inference namespace for sidecar injection
kubectl create namespace ai-inference
kubectl label namespace ai-inference istio-injection=enabled

# Verify installation
istioctl verify-install
istioctl analyze -n ai-inference

主なパターン

クラスター全体での mTLS 厳格モード

apiVersion: security.istio.io/v1beta1
kind: PeerAuthentication
metadata:
  name: default
  namespace: istio-system
spec:
  mtls:
    mode: STRICT
---
# Namespace-level override if needed for gradual rollout
apiVersion: security.istio.io/v1beta1
kind: PeerAuthentication
metadata:
  name: ai-inference-mtls
  namespace: ai-inference
spec:
  mtls:
    mode: STRICT
  portLevelMtls:
    # gRPC inference port
    8081:
      mode: STRICT
    # Prometheus metrics port - allow plaintext scraping
    9090:
      mode: PERMISSIVE

サービスアカウントごとの AuthorizationPolicy

apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: model-server-access
  namespace: ai-inference
spec:
  selector:
    matchLabels:
      app: model-server
  action: ALLOW
  rules:
  - from:
    - source:
        principals:
        - "cluster.local/ns/ai-inference/sa/api-gateway"
        - "cluster.local/ns/ai-inference/sa/orchestrator"
    to:
    - operation:
        methods: ["POST"]
        paths: ["/v1/predict", "/v1/embeddings", "/v2/models/*/infer"]
---
apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: deny-external-to-retriever
  namespace: ai-inference
spec:
  selector:
    matchLabels:
      app: vector-retriever
  action: DENY
  rules:
  - from:
    - source:
        notNamespaces: ["ai-inference"]

承認済みモデルエンドポイント向けの Egress ポリシー

apiVersion: networking.istio.io/v1alpha3
kind: ServiceEntry
metadata:
  name: openai-api
  namespace: ai-inference
spec:
  hosts:
  - api.openai.com
  ports:
  - number: 443
    name: https
    protocol: TLS
  resolution: DNS
  location: MESH_EXTERNAL
---
apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: openai-api-tls
  namespace: ai-inference
spec:
  host: api.openai.com
  trafficPolicy:
    tls:
      mode: SIMPLE
    connectionPool:
      http:
        h2UpgradePolicy: UPGRADE
      tcp:
        maxConnections: 50
---
apiVersion: security.istio.io/v1beta1
kind: AuthorizationPolicy
metadata:
  name: restrict-egress
  namespace: ai-inference
spec:
  action: ALLOW
  rules:
  - to:
    - operation:
        hosts:
        - "api.openai.com"
        - "models.anthropic.com"
        - "*.blob.core.windows.net"

トラフィック管理

A/B モデルテスト向けの VirtualService

apiVersion: networking.istio.io/v1alpha3
kind: VirtualService
metadata:
  name: model-server
  namespace: ai-inference
spec:
  hosts:
  - model-server
  http:
  # Route by header for explicit model version selection
  - match:
    - headers:
        x-model-version:
          exact: "v2-experimental"
    route:
    - destination:
        host: model-server
        subset: v2-experimental
    timeout: 120s
  # Route by header for A/B test cohort
  - match:
    - headers:
        x-ab-cohort:
          exact: "treatment"
    route:
    - destination:
        host: model-server
        subset: v2-experimental
      weight: 100
    timeout: 120s
  # Default traffic split: 90/10 canary
  - route:
    - destination:
        host: model-server
        subset: v1-stable
      weight: 90
    - destination:
        host: model-server
        subset: v2-experimental
      weight: 10
    timeout: 60s
    retries:
      attempts: 2
      perTryTimeout: 30s
      retryOn: unavailable,resource-exhausted

サブセットを使用する DestinationRule

apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: model-server
  namespace: ai-inference
spec:
  host: model-server
  trafficPolicy:
    connectionPool:
      http:
        h2UpgradePolicy: UPGRADE
        maxRequestsPerConnection: 100
      tcp:
        maxConnections: 200
        connectTimeout: 5s
    loadBalancer:
      simple: LEAST_REQUEST
  subsets:
  - name: v1-stable
    labels:
      version: v1
    trafficPolicy:
      connectionPool:
        http:
          maxRequestsPerConnection: 50
  - name: v2-experimental
    labels:
      version: v2
    trafficPolicy:
      connectionPool:
        http:
          maxRequestsPerConnection: 20

推論バックエンドのサーキットブレーカー

apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: model-server-circuit-breaker
  namespace: ai-inference
spec:
  host: model-server
  trafficPolicy:
    connectionPool:
      tcp:
        maxConnections: 100
        connectTimeout: 10s
      http:
        http1MaxPendingRequests: 50
        http2MaxRequests: 200
        maxRequestsPerConnection: 10
        maxRetries: 3
    outlierDetection:
      consecutive5xxErrors: 3
      interval: 15s
      baseEjectionTime: 30s
      maxEjectionPercent: 50
      minHealthPercent: 30
      splitExternalLocalOriginErrors: true
---
# Separate circuit breaker for the vector retriever
apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: vector-retriever-circuit-breaker
  namespace: ai-inference
spec:
  host: vector-retriever
  trafficPolicy:
    connectionPool:
      tcp:
        maxConnections: 300
      http:
        http1MaxPendingRequests: 200
        http2MaxRequests: 500
    outlierDetection:
      consecutive5xxErrors: 5
      interval: 10s
      baseEjectionTime: 15s
      maxEjectionPercent: 30

ストリーミングリクエスト向けのリトライ予算

apiVersion: networking.istio.io/v1alpha3
kind: VirtualService
metadata:
  name: streaming-inference
  namespace: ai-inference
spec:
  hosts:
  - model-server
  http:
  # Streaming endpoint: no retries, long timeout
  - match:
    - uri:
        prefix: /v1/stream
    route:
    - destination:
        host: model-server
        subset: v1-stable
    timeout: 300s
    retries:
      attempts: 0
  # Embeddings endpoint: safe to retry, short timeout
  - match:
    - uri:
        prefix: /v1/embeddings
    route:
    - destination:
        host: model-server
        subset: v1-stable
    timeout: 15s
    retries:
      attempts: 3
      perTryTimeout: 5s
      retryOn: 5xx,reset,connect-failure,retriable-status-codes

レジリエンス

ローカリティを考慮したルーティング

apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: model-server-locality
  namespace: ai-inference
spec:
  host: model-server
  trafficPolicy:
    loadBalancer:
      localityLbSetting:
        enabled: true
        distribute:
        - from: "us-east-1/us-east-1a/*"
          to:
            "us-east-1/us-east-1a/*": 80
            "us-east-1/us-east-1b/*": 20
        failover:
        - from: us-east-1
          to: us-west-2
    outlierDetection:
      consecutive5xxErrors: 3
      interval: 10s
      baseEjectionTime: 30s

可観測性

# Telemetry resource for custom metrics on inference services
apiVersion: telemetry.istio.io/v1alpha1
kind: Telemetry
metadata:
  name: inference-telemetry
  namespace: ai-inference
spec:
  metrics:
  - providers:
    - name: prometheus
    overrides:
    - match:
        metric: REQUEST_DURATION
        mode: CLIENT_AND_SERVER
      tagOverrides:
        model_name:
          operation: UPSERT
          value: "request.headers['x-model-name']"
        tenant_id:
          operation: UPSERT
          value: "request.headers['x-tenant-id']"
  tracing:
  - providers:
    - name: zipkin
    randomSamplingPercentage: 10.0

Kiali ダッシュボードの確認

# Port-forward Kiali
kubectl port-forward svc/kiali -n istio-system 20001:20001 &

# Verify mesh health via API
curl -s http://localhost:20001/kiali/api/namespaces/ai-inference/health | jq .

# Check proxy sync status
istioctl proxy-status -n ai-inference

# Debug a specific pod sidecar config
istioctl proxy-config routes deploy/model-server -n ai-inference -o json
istioctl proxy-config cluster deploy/model-server -n ai-inference

避けるべき落とし穴

• タイムアウトを短く設定しすぎるとストリーミングレスポンスが中断されます。生成エンドポイントでは 300s 以上に設定してください
• 一律のリトライは高価な生成呼び出しを増幅させます。冪等でないルートではリトライを無効にしてください
• テナント向けサービスと内部サービスの間に、アイデンティティの境界を設けない
• 厳格な mTLS からヘルスチェックポートとメトリクスポートを除外し忘れる
• 小規模なプールでアウトライアーの排除を厳しく設定しすぎる（maxEjectionPercent が高すぎる）
• holdApplicationUntilProxyStarts を使用しないと、起動時に競合状態が発生する

関連スキル

• サービスメッシュ - メッシュの基本概念
• llm-gateway - 南北方向の API ゲートウェイ制御
• opentelemetry - エンドツーエンドのトレースとメトリクス