metaflow-mage

Run any Metaflow flow as a Mage pipeline without rewriting your steps.

The problem

Mage is a modern data pipeline tool with a rich UI and flexible block model, but it has no native support for Metaflow's step graph, artifact store, or retry semantics. Teams that want Mage's notebook-style visibility must either re-implement their Metaflow flows as Mage blocks from scratch or give up Metaflow's lineage, versioning, and @retry behavior. There is no ready-made compiler that maps a Metaflow DAG to Mage blocks while keeping the full Metaflow runtime intact.

Quick start

pip install metaflow-mage
python flow.py mage create --mage-host http://localhost:6789
python flow.py mage trigger --mage-host http://localhost:6789
# Compiling HelloFlow to Mage pipeline...
# Pipeline hello_flow deployed successfully.
# Pipeline run started: http://localhost:6789/pipelines/hello_flow/runs/1

Install

pip install metaflow-mage

From source:

git clone https://github.com/npow/metaflow-mage
cd metaflow-mage
pip install -e .

Usage

Deploy and run in one step:

python flow.py mage run \
  --mage-host http://localhost:6789 \
  --mage-project metaflow_project
# Compiling HelloFlow to Mage pipeline...
# Pipeline hello_flow deployed successfully.
# Triggering pipeline run...
# Pipeline run started: http://localhost:6789/pipelines/hello_flow/runs/1
# Pipeline run is running...
# Pipeline run 1 completed successfully.

Deploy once, trigger many times with parameters:

python flow.py mage create --mage-host http://localhost:6789
python flow.py mage trigger --mage-host http://localhost:6789 \
  --run-param message=hello --run-param n=10

Programmatic API:

from metaflow import Deployer

with Deployer("flow.py") as d:
    df = d.mage().create(
        mage_host="http://localhost:6789",
        mage_project="metaflow_project",
    )
    run = df.trigger(message="hello")
    print(run.status)       # RUNNING / SUCCEEDED / FAILED
    print(run.mage_ui)      # http://localhost:6789/pipelines/hello_flow/runs/1
    print(run.run.successful)  # True

Delete a deployed pipeline:

python flow.py mage delete --mage-host http://localhost:6789
# Pipeline hello_flow deleted from Mage.

How it works

Each Metaflow step becomes a Mage Python block. An init block runs first to allocate the Metaflow run ID; subsequent blocks call python flow.py step <step_name> with --run-id, --task-id, and --retry-count threaded through as block outputs. Foreach steps emit a foreach_init block that enumerates items and fans out to per-item body blocks. A Mage API trigger (pipeline schedule) is created or reused on every create, so trigger works immediately without manual Mage UI setup.

Supported graph patterns: linear, branch/join, foreach. @parallel is not supported.

Configuration

CLI option	Environment variable	Default	Description
--mage-host	MAGE_HOST	http://localhost:6789	Mage server base URL
--mage-project	MAGE_PROJECT	metaflow_project	Mage project name
--max-workers	—	10	Max concurrent block runs
--branch	—	—	@project branch name
--production	—	false	Deploy to the production project branch
--name	—	derived from class name	Override the Mage pipeline UUID

Development

git clone https://github.com/npow/metaflow-mage
cd metaflow-mage
pip install -e ".[dev]"
pytest tests/

Integration tests require a running Mage instance:

docker run -it -p 6789:6789 mageai/mageai mage start metaflow_project
pytest tests/ -m integration

License

Apache 2.0. See LICENSE.