Airflow + Ray Demo

🧪 Experimental Version

This provider is an experimental alpha containing necessary components to orchestrate and schedule Ray tasks using Airflow. It is actively maintained and being developed to bring production-ready workflows to Ray using Airflow. This release contains everything needed to begin building these workflows using the Airflow Taskflow API.

Copy
Current Release: 0.2.1

Requirements

Visit the Ray Project page for more info on Ray.

⚠️ The server version and client version (build) of Ray MUST be the same.

Copy
- Python Version >= 3.7
- Airflow Version >= 2.0.0
- Ray Version == 1.3.0
- Filelock >= 3.0.0

Modules

• Ray XCom Backend: Custom XCom backend to assist operators in moving data between tasks using the Ray API with its internal Plasma store, thereby allowing for in-memory distributed processing and handling of large data objects.
• Ray Hook: Extension of Http hook that uses the Ray client to provide connections to the Ray Server.
• Ray Decorator: Task decorator to be used with the task flow API, combining wrapping the existing airflow @task decorate with Ray.remote functionality, thereby executing each task on the Ray cluster.

Configuration and Usage

1. Install the astro-cli. This project was made using the astro dev init command, but that has already been done for you.

2. In your Airflow Dockerfile, your docker file should look something like this:

   Copy
   FROM quay.io/astronomer/ap-airflow:2.0.2-1-buster-onbuild
   USER root
   RUN pip uninstall astronomer-airflow-version-check -y
   USER astro
   ENV AIRFLOW__CORE__XCOM_BACKEND=ray_provider.xcom.ray_backend.RayBackend

   Check ap-airflow version, if unsure, change to ap-airflow:latest-onbuild. Please also feel free to add any pip packages to the requirements.txt but note that these packages will only exist in airflow, and will need to be installed on Ray separately.

3. Configure Ray Locally. To run Ray locally, you'll need a minimum 6GB of free memory. To start, in your environment with Ray installed, run:

   Copy
   (venv)$ Ray start --num-cpus=8 --object-store-memory=7000000000 --head

   If you have extra resources, you can bump the memory up.

   You should now be able to open the Ray dashboard at http://127.0.0.1:8265/.

4. Start your Airflow environment and open the UI. If you have installed the astro CLI, you can do this by running astro dev start.

5. In the Airflow UI, add an Airflow Pool with the following:

   Copy
   Pool (name): ray_worker_pool
   Slots: 25

6. If you are running Ray locally, get your IP address by visiting ipv4.icanhazip.com

7. In the Airflow UI, add an Airflow Connection with the following:

   Copy
   Conn Id: ray_cluster_connection
   Conn Type: HTTP
   Host: Cluster IP Address, with basic Auth params if needed
   Port: 10001

8. In your Airflow DAG Python file, you must include the following in your default_args dictionary:

   Copy
   from ray_provider.xcom.ray_backend import RayBackend
   .
   .
   .
   default_args = {
       'on_success_callback': RayBackend.on_success_callback,
       'on_failure_callback': RayBackend.on_failure_callback,
       .
       .
       .
   }
   @dag(
       default_args=default_args,
       .
       .
   )
   def ray_example_dag():
       # do stuff

9. Using the taskflow API, your airflow task should now use the @ray_task decorator for any Ray task and add the ray_conn_id, parameter as task_args, like:

   Copy
   from ray_provider.decorators import ray_task
   
   default_args = {
       'on_success_callback': RayBackend.on_success_callback,
       'on_failure_callback': RayBackend.on_failure_callback,
       .
       .
       .
   }
   task_args = {"ray_conn_id": "ray_cluster_connection"}
   .
   .
   .
   @dag(
       default_args=default_args,
       .
       .
   )
   def ray_example_dag():
   
       @ray_task(**task_args)
       def sum_cols(df: pd.DataFrame) -> pd.DataFrame:
           return pd.DataFrame(df.sum()).T

Project Contributors and Maintainers

This project is built in collaboration between Astronomer and Anyscale, with active contributions from:

• Pete DeJoy
• Daniel Imberman
• Rob Deeb
• Richard Liaw
• Charles Greer
• Will Drevo