MLOps — ML Operations — applies DevOps principles to machine learning. The problem: a model that works well in a data scientist's notebook behaves differently in production. The "deployment gap" is the biggest issue in the field. 87% of ML projects never reach production.
The MLOps stack: data versioning (DVC), experiment tracking (MLflow, Weights & Biases), model registry, CI/CD pipeline (GitHub Actions), model serving (FastAPI, TorchServe, BentoML), monitoring (data drift, model performance degradation). The feature store manages real-time and batch features. A/B testing in production lets you compare a new model against the old one.
Our courses on this topic