Coinbase's ML Training Evolution: From Sagemaker to Ray
summary
Coinbase replaces Sagemaker with Ray and rebuilds training as small wheel/zip deployments plus Ray Data and a long‑running cluster, cutting PR→run iteration from hours to seconds, reducing last‑mile transform from 120 to 15 minutes, and enabling TB and 50x larger tabular jobs including terabyte-level datasets (reads of 200 GB in 2 minutes via data‑sharing), adding shared cached environments, HPO resource sharing and achieving about 20% cost reduction in instance usage.
AI-written summary · about
chapters (10)
- 0:00Introduction and background
- 1:06Machine learning use cases
- 2:48Migration overview
- 3:37Improving iteration speed
- 6:42Optimizing training workflows
- 10:18Scaling data processing
- 12:40Distributed data and infrastructure
- 16:49Efficiency and cost management
- 17:45Migration results and process
- 19:14Challenges and future outlook
related talks
16mFrom Self-Managed Ray to Anyscale Services: Migrating ML Serving
44mPanel: ML Platform on Ray
23mHeterogeneous Pipelines as the Unlock for ML at Scale
24mBuilding Complex Data Analytics Pipelines with Ray
21mLeveraging Ray to Enhance Machine Learning Models in Finance
20mThe Quick Journey to Using Ray: How We Implement Ray and Anyscale to Speed up our ML Processes
16mScaling Massive Transformer Training with KubeRay
05mInterview