Sponsored Session: MLOps: Why DevOps Solutions Fall Short in the Machine Learning World
Talkat Open Source Summit Europe 2023 · on Tuesday, 2023-09-19· at 15:40· Room 0D-2-0D-3 (Floor 0), Abandoibarra Etorb., 4, 48011 Bilbao, Bizkaia, Spain·
summary
Argues that DevOps tooling misses ML because people and artifacts differ: data scientists (not software engineers) prefer Jupyter-first workflows, need production data, and models encode implicit logic. Gives concrete failures: a platform replaced Uzi pipelines with microservices and adoption stalled for 300 data scientists; Kubeflow lacked a model registry. Details GitLab work: native Jupyter diffs (released 14.5), a built-in MLflow-compatible model experiments feature (saves artifacts, ties to CI), upcoming model registry, and limits like 10 GB SaaS artifact size and GPU-enabled runners.
AI-written summary · about
chapters (18)
- 0:00MLOps introduction
- 1:05What is MLOps vs DevOps
- 2:37DevOps limits for data science
- 7:10Why DevOps falls short for machine learning
- 11:00ML development differences
- 16:18GitLab’s MLOps approach
- 19:06Unified MLOps vision
- 22:17Jupyter notebook reviews in GitLab
- 23:00Model experiment tracking in GitLab
- 24:23MLflow integration in GitLab
- 26:30Model registry in GitLab
- 27:57LLMOps challenges
- 30:05MLOps summary
- 32:21Q&A on data versioning
- 34:04Q&A on model registry
- 35:27Q&A on model deployment
- 36:35Q&A on LLM deployment
- 38:23Q&A on model explainability
related talks
41mMLOps: Why DevOps Solutions Fall Short in the Machine Learning...
14mWhy Your DevOps Skills Are Your MLOps
36mIs the MLOps tooling the same as the DevOps tooling?
44mDevOps for Machine Learning - why is it different?
01mMLOps: Applying DevOps to ML
41mTHE STATE OF MLOPS - MACHINE LEARNING IN PRODUCTION AT ENTERPRISE SCALE
13mMLOps at Scale
47mMachine Learning in Production: An Intro to MLOps.