Unravelling Hidden Technical Debt in ML: A Pythonic Approach to Robust Systems
summary
Claims six core ML challenges illustrated by a churn‑prediction case and shows how quick notebook→Lambda deployment produces training‑serving skew and brittle configs that break replication. It builds containerized pipelines using Docker and Flyte, keeps artifacts in an S3 model registry, uses Airflow for complementary orchestration, tracks code with Git, runs containers on Kubernetes managed by Terraform, adds monitoring and a feature store and uses autoscaling so the system becomes observable and far easier to maintain.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:47Early ML project mistakes
- 4:10Hidden technical debt signs
- 6:58ML debt is harder
- 8:02Robust ML system components
- 10:53Reliable data platform
- 12:17Automating with Prefect
- 14:25Model registry and versioning
- 15:28Production serving and config
- 16:33Monitoring and observability
- 17:39Kubernetes and infrastructure as code
- 18:20Feature stores for consistency
- 19:44Ethical considerations and explainability
- 21:27Key lessons and shortcuts
- 23:14Summary and resources
- 24:39Q&A
related talks
42mDeploying ML models and all the things that go wrong
39mApplying deployment oriented mindset for building Machine Learning models
31m🤑 Paying Off Technical Debt in the Development of Machine Learning Apps
21mMastering the DevOps Machine Learning Pipeline
33mCompliant Data Management and Machine Learning on Kubernetes
37mClean Architecture: How to Structure Your ML Projects to Reduce Technical Debt
24mFrom Model to Production: Automating the ML Lifecycle with CI/CD
30mChallenges of machine learning: The new ML lifecycle