Migrating from Airflow to Flyte

Important

Many Airflow operators and sensors have been tested on Flyte, but some may not work as expected. If you encounter any issues, please file an issue or reach out to the Flyte community on Slack.

Flyte can compile Airflow tasks into Flyte tasks without changing code, which allows you to migrate your Airflow DAGs to Flyte with minimal effort.

In addition to migration capabilities, Flyte users can seamlessly integrate Airflow tasks into their workflows, leveraging the ecosystem of Airflow operators and sensors. By combining the robust Airflow ecosystem with Flyte’s capabilities such as scalability, versioning, and reproducibility, users can run more complex data and machine learning workflows with ease. For more information, see the Airflow agent documentation.

For current Flyte users

Even if you’re already using Flyte and have no intentions of migrating from Airflow, you can still incorporate Airflow tasks into your Flyte workflows. For instance, Airflow offers support for Google Cloud Dataproc Operators, facilitating the execution of Spark jobs on Google Cloud Dataproc clusters. Rather than developing a custom plugin in Flyte, you can seamlessly integrate Airflow’s Dataproc Operators into your Flyte workflows to execute Spark jobs.

Prerequisites

  • Install flytekitplugins-airflow in your Python environment.

  • Enable an Airflow agent in your Flyte cluster.

Steps

1. Define your Airflow tasks in a Flyte workflow

Flytekit compiles Airflow tasks into Flyte tasks, so you can use any Airflow sensor or operator in a Flyte workflow:

from flytekit import task, workflow
from airflow.sensors.filesystem import FileSensor

@task
def say_hello() -> str:
    return "Hello, World!"

@workflow
def airflow_wf():
    flyte_task = say_hello()
    airflow_task = FileSensor(task_id="sensor", filepath="/")
    airflow_task >> flyte_task

if __name__ == "__main__":
    print(f"Running airflow_wf() {airflow_wf()}")

2. Test your workflow locally

Note

Before running your workflow locally, you must configure the Airflow connection by setting the AIRFLOW_CONN_{CONN_ID} environment variable. For example,

export AIRFLOW_CONN_MY_PROD_DATABASE='my-conn-type://login:password@host:port/schema?param1=val1&param2=val2'

Although Airflow doesn’t support local execution, you can run your workflow that contains Airflow tasks locally, which is helpful for testing and debugging your tasks before moving to production.

AIRFLOW_CONN_FS_DEFAULT="/" pyflyte run workflows.py airflow_wf

Warning

Some Airflow operators may require certain permissions to execute. For instance, DataprocCreateClusterOperator requires the dataproc.clusters.create permission. When running Airflow tasks locally, you may need to set the necessary permissions locally for the task to execute successfully.

3. Move your workflow to production

Note

In production, we recommend storing connections in a secrets backend. Make sure the agent pod has the right permission (IAM role) to access the secret from the external secrets backend.

After you have tested your workflow locally, you can execute it on a Flyte cluster using the --remote flag. In this case, Flyte creates a pod in the Kubernetes cluster to run the say_hello task, and then runs your Airflow BashOperator task on the Airflow agent.

pyflyte run --remote workflows.py airflow_wf