Applying deployment oriented mindset for building Machine Learning models
summary
Claims hands‑on experience building two deployment services and improving their architecture, with concrete tools and fixes: use of SageMaker, Docker/Kubernetes, DVC and MLflow, plus a custom notebook-friendly versioning module. Describes reproducible failure modes—model response inconsistency, feature calculation mismatches, remote DB latency and scraper rate limits—and practical countermeasures: save model iterations, unit tests for feature logic, timing measurements, refactoring and automated pipelines for multi‑client deployment. Useful if wanting actionable deployment practices.
AI-written summary · about
chapters (37)
- 0:00Introduction
- 0:28Agenda
- 1:37Research vs Deployment
- 3:06Model wrapper
- 4:36Model response inconsistency
- 7:18Features calculation
- 8:08Problems
- 8:19Early stages
- 9:36Deployment
- 11:46Data sources
- 13:22Communication
- 13:42Data Understanding
- 15:16Refactoring
- 15:40Testing
- 17:25Readability
- 18:30Recommended books
- 18:51Test model response
- 20:17Pipelines
- 22:23Document everything
- 22:57Avoid writing documentation
- 23:48What deployment oriented mindset gives you
- 24:46Distribution in your team
- 25:32Thank you
- 25:51Questions
- 26:11SageMaker
- 27:24Automated pipelines
- 28:06Integration
- 29:26Test coverage
- 31:22Unit tests
- 32:20How can we avoid it
- 33:14The only solution
- 34:06The best wave
- 35:02Why Jupiter
- 35:17Why Data Science
- 36:13Versioning
- 36:50Jupiter Notebooks
- 37:44Tradeoff between super cool features and deployment complexity
related talks
42mDeploying ML models and all the things that go wrong
34mDeploying a Machine Learning System That Only Runs Locally
59mDevOps for Machine Learning Projects
57mDevOps for Machine Learning
56mDevOps for Machine Learning
29mDeploying ML Models and the Things That Go Wrong
57mManaging Machine Learning Lifecycle with MLflow
28mUnravelling Hidden Technical Debt in ML: A Pythonic Approach to Robust Systems