Building scalable end-to-end deep learning pipelines in the cloud
summary
Describes a concrete pipeline pattern that uses AWS Batch, AWS Fargate, Amazon SageMaker, AWS Lambda, AWS Step Functions where Step Functions orchestrates five staged jobs, Batch runs heavy CPU/GPU tasks, SageMaker handles training with hyperparameter tuning and spot checkpointing, Fargate hosts long CPU jobs, and Lambda runs short parallel tasks. A demo repository deploys the workflow and shows Batch job logs and completed steps. Notes practical thresholds: 10,000 concurrent Lambdas and that Lambda is cost‑effective under ~1M invocations/month depending on latency and idle‑time tradeoffs.
AI-written summary · about
chapters (13)
- 0:00Introduction and overview
- 2:30What is serverless computing
- 5:20Serverless compute options compared
- 8:10CPU vs GPU for ML workloads
- 12:24Cloud native orchestrators
- 16:59Designing the deep learning pipeline
- 19:29Model training patterns
- 23:44Model inference architecture
- 27:33Serverless inference frameworks
- 30:24When serverless inference works
- 32:13Key takeaways and summary
- 32:54Live demo of pipeline deployment
- 38:45Q&A and resources
related talks
34mBuilding scalable end-to-end deep learning pipelines in the cloud
12mLeveraging AWS and Kubernetes for Scalable AI/ML Workloads in the Cloud
33mBuilding & Managing MLOps Pipelines in the Cloud
34mDeploying a Machine Learning System That Only Runs Locally
25mDesigning an AWS Well Architected Framework for AI/ML
41mBuilding Serverless Machine Learning Models in the Cloud
34mLat Designing and Building Complex Machine Learning....
44mPragmatic Machine Learning in the Cloud