MLOps и Model Serving

MLOps — DevOps практики для ML. Автоматизация lifecycle от experiments до production.

Ключевые компоненты

Experiment tracking: MLflow, Weights & Biases. Model registry: versioning, staging. Feature store: reusable features (Feast, Tecton). Data versioning: DVC, LakeFS. CI/CD для models: testing, deployment. Monitoring в production. Automated retraining pipelines.

Serving patterns

Batch prediction: nightly scoring в data warehouse. Real-time API: low latency (<100ms). Streaming: Kafka Streams + model. Edge deployment: mobile, IoT. Hybrid: batch features + real-time inference. Каждый паттерн для different use case.

Tools

Kubernetes: unified orchestration. Seldon, KServe: model serving на K8s. BentoML для packaging. NVIDIA Triton для GPU. Ray Serve. Managed: SageMaker, Vertex AI, Azure ML. Choice depends on scale, cost, cloud.

Governance

Model cards: documentation каждой модели. Explainability: SHAP, LIME. Fairness audits. Regulatory compliance для automated decisions (GDPR Art. 22). Audit trails. Approval workflows для production deployment. Rollback procedures.