Ramp's data platform team shares how adopting Metaflow transformed their ML development workflow. Before Metaflow, a single model took months to ship due to slow feedback loops, complex dependency management, and heavy platform involvement. After adopting Metaflow alongside AWS Batch and Step Functions, they shipped eight additional models in ten months. The post covers their infrastructure choices (AWS Batch with EC2 over Fargate), how they integrated Metaflow with Airflow via a custom MetaflowOperator, and how data scientists can now largely self-service their ML workflows from prototype to production.
Table of contents
Simplifying Finance with Machine Learning at RampLong Feedback Loops, and Excessive FrictionChoosing MetaflowThe Technical Details of Our SetupWhere We are Today4 Impressions