DevOps for Machine Learning - why is it different?
summary
Claim: machine learning fails in ops because 80–90% never go live and the bulk of effort is data, training and monitoring not code. Demonstrated tools and patterns: Seldon + Kubeflow pipelines, MLflow, Pachyderm, TensorFlow Serving, SageMaker, Prometheus, Elasticsearch. Concrete build: a pickled scikit-learn model packaged for Kubernetes and served with traffic strategies (canary, A/B, shadowing). Operational proof: shadowing plus Elasticsearch request logging and metric tracking enable comparison and rollout decisions, showing the approach works in practice.
AI-written summary · about
chapters (31)
- 0:00Intro
- 2:09MLOps Background
- 3:36Complex Tool Landscape
- 4:26Why So Different?
- 5:05Examples
- 7:09Regression
- 8:00Gradient Descent
- 8:40Data Playgrounds/Exploration
- 9:09Comparing Journeys
- 9:22Dev Build Journey
- 10:13ML Build Journey
- 17:25ML is Different - Key Points
- 18:35Traditional Dev Workflow
- 19:36ML Workflows - Primer
- 20:43ML EZE Workflow Intro
- 24:16What Can Happen
- 25:10Role of MLOps
- 26:07New Territory
- 26:52Training/Experimentation
- 27:38Model Training
- 28:34Kubeflow Pipelines
- 29:37MLFlow Experiments
- 31:41Seldon ML Serving
- 32:44Rollout and Monitoring
- 34:28Canary with Seldon
- 34:52A/B Test With Seldon
- 35:13Seldon Metrics
- 36:11Seldon Request Logging
- 37:22Advanced Topics - Serving
- 38:19Advanced Topics - Governance
- 41:06Summar
related talks
56mDevOps for Machine Learning
57mDevOps for Machine Learning
47mMachine Learning in Production: An Intro to MLOps.
39mContinuous Delivery for Machine Learning — The Future of MLOps
41mSponsored Session: MLOps: Why DevOps Solutions Fall Short in the Machine Learning World
01mMLOps: Applying DevOps to ML
13mMLOps: From AI Research to Production-Ready Systems- 41mMLOps: Why DevOps Solutions Fall Short in the Machine Learning...