Flyte 2: pure Python ML orchestration, open-source backend still loading

What it does

Flyte 2 is a Python framework for orchestrating machine-learning pipelines, models, and agents. You wrap functions in a TaskEnvironment, decorate them with @env.task, and the framework handles container images, dependencies, and execution. It also supports serving models directly via a FastAPIAppEnvironment, so the same codebase can train and host an endpoint.

The interesting bit

The SDK leans into native Python patterns — async/await, type hints, and asyncio.gather — rather than inventing its own workflow DSL. The catch is that the open-source Kubernetes backend is still listed as “coming soon,” which means the self-hosted, multi-node orchestrator this repository promises does not yet exist in the open; production deployments currently route through Union.ai.

Key highlights

• Decorator-based Python API: define tasks and pipelines in code, not configuration files.
• Native asyncio support: compose parallel work with standard Python concurrency instead of framework-specific primitives.
• Integrated model serving: wrap a FastAPI application as an environment and deploy it alongside your training logic.
• Local TUI for debugging workflows before pushing to any remote environment.
• Graduated LF AI & Data Foundation project with a separate, maintained Flyte 1 branch.

Caveats

• The open-source backend is not yet available; this repository currently houses the upcoming Go infrastructure, while the Python SDK lives elsewhere.
• Production-grade, multi-node orchestration today requires Union.ai’s hosted platform, not a self-managed open-source release.
• Flyte 1 remains actively maintained on the master branch, so documentation and community examples may still point to the older version.

Verdict

Worth a look for ML engineers who want to stay in Python and need a structured path from notebook code to containerized tasks. If you require a fully open-source, self-hosted orchestrator right now, you are better off waiting for the backend release or evaluating Flyte 1.