Sunset for the Wild West: Making ML disciplined by default
summary
Claims that dependency chaos makes ML systems insecure and unreproducible, supported by a PyTorch container dependency graph with ~1,500 nodes and ~6,000 edges; proposes prescriptive controls (deny packages newer than a chosen age to avoid fresh supply‑chain exploits) and descriptive observability using OSQuery (about 500 tables of system facts) plus a custom preprocessor that turns SQL in Markdown into validated OSQuery packs, flags missing columns and join‑order pitfalls, materializes change data, and runs as a Kubernetes DaemonSet.
AI-written summary · about
chapters (9)
- 0:00Introduction and disclaimer
- 4:22Creativity in ML systems
- 10:19Reproducibility challenges
- 13:31Reproducibility vs security
- 17:42OSQuery overview
- 24:05OSQuery pain points
- 31:56Tooling for OSQuery
- 35:24Summary and takeaways
- 36:28Q&A
related talks
28mUnravelling Hidden Technical Debt in ML: A Pythonic Approach to Robust Systems
43mHacking and Securing ML Systems and Environments
33mCompliant Data Management and Machine Learning on Kubernetes
25mSecure ML: Automated Security Best Practices in Machine Learning
26mRunning Secure AI/ML Experiment in a Unified MLOPS-DevSecOps Pipeline
48mDeveloper Insights: ML and analytics on src/
21mEnd-to-end secure ML Development
24mChaos and Pain in Machine Learning, and the ‘DevOps for ML Manifesto'