In 2017, Etsy developed an ML platform to help their data science team deliver and iterate on models effectively and efficiently. However, to support a wider range of tools and frameworks, they decided to build a second major version of the ML platform. In this post, Kyle Gallatin and Rob Miles describe the platform’s components for training and prototyping (using Google Vertex AI and Dataflow); model serving (using an internally developed Model Management Service extended with TensorFlow Serving and Seldon Core); and workflow orchestration (with Kubeflow and Vertex AI Pipelines). They discuss how the platform has reduced the time it takes to launch live ML experiments and identify challenges that still remain.