Live demo · always on · no install, no login

Move Airflow tasks with a feature flag

This is a real Apache Airflow 3.x with airflow-provider-openfeature installed, reading a real Flipt backend. A flag moves a subset of tasks to a canary pool, ramps it, and reverts, without editing a DAG.

Change the flag in Flipt on the right, then watch the tasks change pool in Airflow on the left.

Open Airflow →

The platform_pool_canary DAG has 10 tasks asking for the default pool and no flag code. Watch them land in canary_pool.

/dags

Open Flipt →

The real flag backend. Change the rollout on airflow.task.pool or flip the kill switch, then watch Airflow react.

flag admin UI · no login

Try it in about a minute

  1. Open Flipt (no login), find the flag airflow.task.pool, and change its rollout percentage.
  2. Wait ~30–60s for the next DAG parse to pick up the change.
  3. Open Airflow, open platform_pool_canary, click a task, then Instance Details, and read its Pool. The DAG source never changed; the policy set it from the flag.
Provider repo This demo's repo PyPI Docs

The provider is an ordinary pip package that plugs into an Airflow cluster policy and an OpenFeature provider; it does not fork Airflow or patch the scheduler. Flipt is the backend with a UI here; it evaluates over OFREP (HTTP per call), which is fork-safe, so it also backs the live A/B experiment's Python tasks. A headless flagd runs as a reference. The same code works with GO Feature Flag, GrowthBook, LaunchDarkly, or an in-house engine.