Drag-and-Drop Pipelines That Run on HuggingFace Jobs

What it does\nTangle is a web-based orchestrator for any batch workflow with a start and an end. It provides a visual canvas to wire containerized command-line programs into a graph, then executes them locally via Docker or remotely via HuggingFace Jobs and Kubernetes.\n\nThe interesting bit\nThe project treats HuggingFace as more than a model registry: it runs a multi-tenant SaaS on HF Spaces, uses HF authentication, launches compute through HF Jobs, and stores your artifacts in a private HF Dataset repo. If you want isolation, you can duplicate the Space into your own single-tenant instance with its own SQLite database.\n\nKey highlights\n- Drag-and-drop pipeline builder; you can design graphs without creating an account.\n- Execution caching reuses successful—and even currently running—tasks to avoid redundant compute.\n- Components use the ComponentSpec/component.yaml format and are compatible with Kubeflow Pipelines and Google Cloud Vertex AI.\n- Supports arbitrary containers and languages (Python, R, Java, Shell, etc.).\n- Backend separates the API server from the orchestrator, communicating through a SQL database (SQLite or MySQL).\n\nCaveats\n- Running pipelines on the shared HuggingFace instance requires a HuggingFace Pro subscription ($9/month).\n- The README notes that installation recipes are still expanding, so deployment beyond local Docker, Google Cloud Shell, and HuggingFace may require extra legwork.\n\nVerdict\nWorth a look for teams who need reproducible, containerized batch processing but want to skip local environment setup and complex cloud IAM. Skip it if you’re already happy with Airflow, Argo, or a mature in-house scheduler.