MLOps — Machine Learning Operations

MLOps applies DevOps principles to ML — automating model training, deployment, monitoring, and maintenance.

MLOps Lifecycle

Data Pipeline → Training → Evaluation → Deployment → Monitoring
     ↑                                                    │
     └────────────────────────────────────────────────────┘
                         (Feedback Loop)

Tools:
├─ Data: DVC, Feature Store, Great Expectations
├─ Training: MLflow, Weights & Biases, DVC
├─ Deployment: Docker, Kubernetes, Seldon
├─ Monitoring: Evidently, Whylabs, Prometheus
└─ Orchestration: Airflow, Kubeflow, Prefect

Experiment Tracking

import mlflow

# Log experiment
mlflow.log_param("learning_rate", 0.001)
mlflow.log_param("batch_size", 32)
mlflow.log_metric("accuracy", 0.95)
mlflow.log_metric("loss", 0.12)

# Log model
mlflow.sklearn.log_model(model, "model")

Model Deployment

Options:
├─ REST API (FastAPI, Flask)
├─ Serverless (AWS Lambda, GCP Functions)
├─ Container (Docker + Kubernetes)
├─ Edge (ONNX, TFLite)
└─ Managed (SageMaker, Vertex AI)

Model Monitoring

Monitor:
├─ Data drift: Input distribution changes
├─ Model drift: Performance degrades
├─ Prediction drift: Output distribution changes
└─ Latency/throughput

When drift detected → trigger retraining

Key Takeaways

1. MLOps = DevOps for ML
2. Experiment tracking is essential for reproducibility
3. Version control for data, code, and models
4. Automated pipelines reduce manual errors
5. Monitoring detects model degradation
6. A/B testing validates model updates
7. CI/CD for model deployment
8. Feature stores ensure consistent features