TechPrufer
Get started
Spotify
Spotify logoSpotify · Feature trail

Ray Platform for ML

3 posts · Nov 2019 – Jan 2023

What this trail covers

Spotify has introduced a centralized Spotify-Ray platform to democratize ML innovation and support a broader range of ML practitioners, including data and research scientists. This platform leverages Ray to provide flexible, scalable, and accessible infrastructure for ML development, from prototyping to production. It aims to lower the barrier to entry for diverse ML tooling and workflows, including support for novel ML solutions, GPU access, and advanced paradigms like RL and GNNs. The platform also details the evolution of ML infrastructure at Spotify, moving from early custom tooling to standardizing on TensorFlow Extended (TFX) and Kubeflow Pipelines (KFP) for end-to-end ML workflows. This includes managing Kubeflow clusters, providing SDKs and CLIs, and addressing integration challenges between TFX and KFP.

All Spotify trails
  1. Unleashing ML Innovation at Spotify with RayThis post introduces the new centralized Spotify-Ray platform, built on GKE and KubeRay. It details the platform's design principles focused on accessibility, flexibility, availability, and performance. Key features include a CLI for creating, managing, and scaling Ray clusters with pre-installed ML tools, notebook tutorials, VS Code server, and SSH access. The platform aims to support a wider range of ML practitioners beyond ML engineers by enabling easier access to GPUs and distributed compute, accelerating research and prototyping, and productionizing advanced ML paradigms.
  2. Product Lessons from ML Home: Spotify’s One-Stop Shop for Machine LearningThis post details the product lessons learned in building and scaling ML Home, Spotify's internal user interface for its Machine Learning Platform. It describes the evolution from an MVP focused on offline model evaluation to a comprehensive platform integrating with other ML Platform components (Kubeflow, Jukebox, Salem, Klio). Key product lessons include balancing broad vision with iterative strategy, understanding the role of MVPs in validating concepts, and recognizing that the ML Platform's capabilities are ML Home's unique differentiator. The post highlights the importance of metadata services, flexible interfaces (tagging, notes), and information-rich model comparisons. It also discusses scaling efforts, leading to a 200% growth in daily active users and entrenchment in daily workflows.
  3. The Winding Road to Better Machine Learning Infrastructure Through Tensorflow Extended and KubeflowThis post details Spotify's journey in building ML infrastructure, moving from early custom Scala-based tooling to standardizing on TensorFlow Extended (TFX) and Kubeflow Pipelines (KFP). It outlines three iterations: the first focused on connectors for existing Scala data tooling, the second on standardizing on TFX for its Python-centric ML community and robust components, and the third on adopting Kubeflow Pipelines for orchestration, component sharing, and experiment tracking. The post also highlights the challenges and solutions in integrating TFX and KFP, and the operational aspects of managing Kubeflow clusters on GCP, including open-sourcing a Terraform module for GKE cluster setup.