How to deploy machine learning models to production (frequently and safely)
summary
Shows how to move a model from notebook to production with an automated CI/CD pipeline that trains, evaluates, versions artifacts in a GCP bucket model artifacts, and deploys by switching the production pointer. Pipeline runs unit tests plus a metrics-gate test that fails builds below a stated threshold, retrains and stores named artifacts, and supports canary routing that duplicates requests to candidates for live comparison. Claims reduce deployment time from weeks/months to less than 30 minutes and emphasizes rehearsal, rollback and collecting labels for continuous improvement.
AI-written summary · about
chapters (17)
- 0:00Introduction to ML deployment
- 2:30Why deploy models safely
- 7:23Safety enables iterative improvement
- 9:09Continuous delivery principles
- 10:12Feedback loops in CI/CD
- 10:55Challenge one snowflake environments
- 13:05Challenge two testing model changes
- 15:19Challenge three CI pipeline necessity
- 17:17Challenge four model versioning
- 19:02Challenge five automated deployments
- 21:07Challenge six canary releases
- 22:32Challenge seven simple models first
- 24:19Challenge eight traceability and feedback
- 25:46Challenge nine cross-functional teams
- 26:52Challenge ten deployment checklists
- 31:14Conclusion and next steps
- 32:17Q&A session
related talks
35mGetting Machine Learning Models Ready For Production Using Python
57mDevOps for Machine Learning
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
22mFrom experiment to production - a journey of a machine learning model
29mData Science in Production
25mBuilding Scalable MLOps Pipelines That Actually Work
51mSpeeding up deployment of machine learning models
28mLeading machine learning research journeys with happy (into-production) endings