Skip to content

Level 3 · Pipelines & Deployment

Level 3 turns manual, one-off training and deployment steps into a real, automated pipeline: a real training script that runs the same way every time, a real CI check that can block a bad model before it ships, and real deployment patterns (canary, shadow) that limit the blast radius of a model that turns out wrong.

What was run, and what wasn't

This level builds and runs real training scripts, a real local CI pipeline (using real test assertions against a real model's real predictions), and a real containerized model server. A full managed CI/CD platform and a real production-scale canary rollout across many real servers were not run as part of this course; the underlying mechanisms are demonstrated directly and locally instead.

Prerequisites: Levels 1–2.

Modules

  1. Training Pipelines: Notebook to Reproducible Script: a real notebook's logic, turned into a real, repeatable script with identical real output.
  2. Pipeline Orchestration: Steps, Dependencies, Artifacts: a real multi-step pipeline where each step's real output feeds the next.
  3. CI for ML: Testing a Model Before It Ships: a real test that catches a real model regression before deployment.
  4. CD for ML: Automating Deployment of a New Version: a real automated deploy, triggered and verified.
  5. Real-Time vs Batch Inference: Two Deployment Shapes: the same real model, served two genuinely different real ways, measured.
  6. Containerizing a Model for Deployment: a real container running a real model, verified identically to the uncontainerized version.
  7. Canary & Shadow Deployment for Models: a real new model version tested against real traffic before fully replacing the old one.
  8. Feature Stores & Training/Serving Skew: a real, measured case where training-time and serving-time features genuinely disagree.
  9. A/B Testing Models in Production: a real, measured comparison between two real model versions under real traffic.
  10. Project: A Real CI/CD Pipeline for a Model: a complete, real pipeline from commit to deployed model.

After Level 3

Level 4 assumes a model is now live and asks the real, ongoing question: is it still working, and how would you actually know if it wasn't?