MLOps

Status: 🚧 Coming soon — chapters are being written.

MLOps is the discipline of running ML systems in production reliably. It's where data engineering, software engineering, and ML meet. If you've ever shipped a model and then watched it silently rot — this is what you needed to know.

What this section will cover

• The ML lifecycle — research → training → serving → monitoring → retraining
• Experiment tracking — MLflow, Weights & Biases, Neptune
• Data versioning — DVC, LakeFS, Pachyderm
• Model registries and reproducibility
• Serving — REST endpoints, batch inference, streaming, BentoML, KServe, Ray Serve, vLLM for LLMs
• Monitoring — drift, performance decay, fairness, latency, cost
• CI/CD for ML — testing data, models, pipelines
• Feature stores — Feast, Tecton
• LLMOps specifics — prompt versioning, eval pipelines, cost tracking, observability via LangSmith

Currently available — related material

• FastAPI → ML Model Deployment
• FastAPI → Docker Deployment
• LangSmith — observability for LLM apps

A consolidated MLOps track lands next.