GoBubble migrates 21 workloads from GCP to Amazon EKS with zero downtime

GoBubble’s models depend on a continuous supply of features and training data produced by dozens of interdependent pipelines that had grown up inside GCP’s orchestration. Migrating them was the riskiest part of the platform move: every pipeline had to produce byte-equivalent outputs on AWS, on schedule, or model quality would silently degrade — and the whole translation had to land inside the same fixed 12-week window as the rest of the estate, with the moderation product running live throughout.

VeUP re-founded the data plane on AWS-native services. Feature and training data land in an Amazon S3 ML data lake encrypted with SSE-KMS and cost-managed by Intelligent-Tiering; the analytical and feature warehouse is Snowflake on AWS, reached privately over AWS PrivateLink; and the ~85 pipelines were translated onto AWS Glue for ETL, with Snowpark Python carrying the transforms, AWS Step Functions expressing the orchestration logic, and Amazon EventBridge driving schedules and triggers. Translation was governed by an operator-mapping discipline — each legacy pipeline’s behavior mapped explicitly onto its Glue/Step Functions equivalent — and validated the only way that counts: a 14-day feature-data parity-overlap window ran both planes side by side and compared outputs, returning a Pass before the legacy plane was retired. The cutover was zero-downtime, and the platform runs in production on the AWS-native data plane today.