Supercharging the ML and AI Development Experience at Netflix (netflixtechblog.medium.com)

Netflix has added a new development mode called spin to Metaflow 2.19 to make iterative ML/AI development feel as fast and stateful as working in a notebook while keeping code production-ready. Metaflow already treats each @step as a checkpoint that automatically persists instance variables as artifacts; spin takes this further by executing a single @step quickly using the parent step’s in-memory state, skipping the full run metadata so iterations are near-instant. That lets engineers iterate, debug and build visualizations (Metaflow Cards) interactively, then promote the same flow to production orchestrators like Maestro or Argo and scale on AWS Batch, Titus or Kubernetes. Technically, spin distinguishes itself from run and resume by defaulting to ephemeral, non-versioned execution (but can persist outputs to a directory with --persist), accepting injected inputs via an ARTIFACTS module, and exposing results through inspect_spin and the Runner API for unit tests. It integrates with editors (VS Code/Cursor) via hotkeys and supports programmatic usage for agent-driven development—shown in examples where Claude Code uses spin to surface and fix data issues. The feature accelerates the inner loop of model development, improves reproducibility of step-level logic, and enables faster, testable iteration without sacrificing production hygiene.