CourseAI / ML Engineer: The Complete Course · Module 10: MLOps · part 62 of 93
Part 62 · Module 10: MLOps

Lesson 2: Shipping a model without betting the warnings on it

29 min read·
Module 10 · Lesson 2 · About 55 minutes · Needs: the registry, aliases and gates (lesson 1), CI basics and a workflow file (Module 2 lesson 2), significance and paired comparisons (Module 6 lesson 1), the scoring service and prepare (Module 9) and the Kestrel data (course-guide/06).

The situation

Week 10, Wednesday. The service still answers with Module 9's lgbm-2; lesson 1's recipe beats it on the gate weeks. Rhys will not flip 420,000 predictions a day on one offline number: "Show me it behaves on live traffic, and that you can undo it in seconds."

The running example: version 2 (lesson 1's recipe) replacing version 1 (lgbm-2) in this notebook's registry. On 7 to 20 September version 2 scores PR-AUC 0.4109 against 0.3981.

WHERE THIS FITS The Kestrel system this lesson SHARED SERVICES Depot scans parcels, routes Ingestion nightly batch Validation Pandera checks Data lake Parquet, DuckDB Experiments MLflow runs Training LightGBM, Optuna Online store Feast, Redis Features point in time Gates vs the champion Registry staged versions Serving API FastAPI, ONNX Consumers vans, customers Retraining trigger, budget Alerting on call Monitoring drift, Evidently Late labels a day later Orchestrator Prefect flows Versions DVC, Git CI/CD GitHub Actions Containers Docker images
This lesson ships a registered model to the serving API and its consumers, through CI and containers (blue); the rest of Kestrel is grey for now.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.

Buy this course · ₹399See what is inside

Or open everything with a plan