Migrating enterprise ML workloads from Databricks to AWS for large scale ML

This article details Kargo's migration of enterprise ML workloads from Databricks to AWS, achieving significant improvements in cost, scalability, and operational efficiency.

• Replaced Delta Lake ETL with AWS Glue and Apache Iceberg for ACID transactions and schema evolution
• Consolidated scattered modeling logic into containerized Python packages deployed via Amazon ECR
• Implemented SageMaker Pipelines for end-to-end orchestration with deterministic artifact versioning
• Achieved 40% cost reduction through serverless AWS Glue and Athena replacing persistent clusters
• Improved pipeline execution speed 3-5x through parallel SageMaker pipeline execution
• Decoupled real-time inference serving from training using sidecar containers for zero-downtime updates
• Standardized observability via Amazon CloudWatch for unified monitoring across all components
• Maintained byte-for-byte output parity with original Databricks pipelines for production safety

The migration demonstrates how thoughtful re-architecture—rather than lift-and-shift—enables scalable ML platforms supporting both offline optimization and real-time inference at advertising scale.