Facets of Reproducible research on Kubernetes
summary
Kubernetes plus CI/CD and container images make environments reproducible, but real reproducibility requires concrete steps: use binder or declarative images for immutable runs, prefer Kubeflow notebooks for scalable development with explicit volume and image pairing, freeze libraries with pip freeze and record who built numpy (BLAS/Intel/OpenBLAS) and exact versions, and seed RNGs instead of relying on globals (the slide demonstrates 1024-sample streams). Also instrument CI to auto-generate model services, track metrics for concept drift, and beware that excessive reproducibility can cause hidden overfitting from repeated experiment tuning.
AI-written summary · about
chapters (17)
- 0:00Introduction to reproducibility
- 1:12Containers and Kubernetes benefits
- 3:57Why containers aren't enough
- 5:02Causes of irreproducibility
- 5:23Human assumptions in notebooks
- 7:28Library version pitfalls
- 9:57Python dependency issues
- 12:23Library behavior changes
- 13:45Soft dependency challenges
- 15:08Random number generator issues
- 18:16Implicit randomness in libraries
- 18:58Distributed randomness problems
- 19:59Hashing environment differences
- 22:09Data reproducibility barriers
- 23:12Avoiding human dependency
- 24:37Overfitting from reproducibility
- 27:51Summary and closing
related talks
30mRealizing End to End Reproducible Machine Learning on Kubernetes
33mCompliant Data Management and Machine Learning on Kubernetes
20mTrust but Verify
31mKubernetes and MLOps for Scalable and Reproducible Generative AI
31mImage Provenance and Security in Kubernetes
26mKubernetes and MLOps for Scalable and Reproducible Generative AI
51mMachine Learning using Kubeflow and Kubernetes
01h 00mJava on Kubernetes What I Wish I Knew First