Configurable Metaflow (netflixtechblog.com)

Netflix’s Metaflow (v2.13) now includes a first-class Config object to make flows configurable without changing code: configs are resolved at deployment time, persisted as artifacts, and can be referenced anywhere (including decorators like @resources and @schedule) via a convenient dot-syntax. The Config feature answers a long-standing community need to vary CPU/memory, schedules, dependencies, and other flow-level settings per deployment or experiment, enabling teams to run many closely related flow variants (e.g., different data, algorithms, or resource footprints) while keeping the flow structure intact. Technically, Metaflow distinguishes artifacts (persisted at task end), Parameters (resolved at run start), and Configs (resolved at deployment), so Configs can influence decorator behavior and set parameter defaults. Configs support custom parsers (TOML, Pydantic validators, Hydra/OmegaConf, or even remote fetchers), work seamlessly with remote execution, Runner/Deployer APIs, and are versioned and queryable via the Client API. This unlocks advanced workflows—Hydra-driven or Metaboost-managed experiment sweeps that change resources or dependencies per run—making large-scale reproducible experimentation and production deployments far lower-boilerplate and more flexible for ML/AI teams.