Deploying a Machine Learning System That Only Runs Locally
summary
Describes automating deployment of data scientist notebooks using AWS Step Functions vs SageMaker Pipelines, versioning data and models on S3, storing embeddings in scalable storage, and using Dockerfile templating for reproducibility. Compares cost and latency of serverless versus provisioned SageMaker endpoints, shows turning training/eval instances on/off while keeping inference always on, identifies wasted resources (e.g., duplicated variables doubling memory), and proposes code reviews plus drift metrics to decide when to rebuild models.
AI-written summary · about
chapters (11)
- 0:00Local ML intro
- 2:11Classic web actors
- 4:40ML lifecycle
- 10:10Data scientist vs cloud roles
- 16:18Automating notebooks
- 21:48Team workflow
- 23:12Resource management & versioning
- 26:47Embedding tools & storage
- 28:55Step Functions vs SageMaker
- 31:04Inference optimization
- 33:21Q&A
related talks
12mLeveraging AWS and Kubernetes for Scalable AI/ML Workloads in the Cloud
39mApplying deployment oriented mindset for building Machine Learning models
33mBuild, train, and deploy ML models using Amazon SageMaker
31mProductizing Machine Learning over Big Data with AWS tools
37mStreamlining Machine Learning Operations
25mDesigning an AWS Well Architected Framework for AI/ML
34mBuilding scalable end-to-end deep learning pipelines in the cloud
45mDeploying ML Models to Production with Amazon SageMaker