Airflow Orchestrator

Orchestrating your pipelines to run on Airflow.

ZenML pipelines can be executed natively as Airflow DAGs. This brings together the power of the Airflow orchestration with the ML-specific benefits of ZenML pipelines. Each ZenML step runs in a separate Docker container which is scheduled and started using Airflow.

If you're going to use a remote deployment of Airflow, you'll also need a remote ZenML deployment.

When to use it

You should use the Airflow orchestrator if

• you're looking for a proven production-grade orchestrator.

• you're already using Airflow.

• you want to run your pipelines locally.

• you're willing to deploy and maintain Airflow.

How to deploy it

The Airflow orchestrator can be used to run pipelines locally as well as remotely. In the local case, no additional setup is necessary.

There are many options to use a deployed Airflow server:

• Use the ZenML GCP Terraform module which includes a Google Cloud Composer component.

• Use a managed deployment of Airflow such as Google Cloud Composer , Amazon MWAA, or Astronomer.

• Deploy Airflow manually. Check out the official Airflow docs for more information.

If you're not using the ZenML GCP Terraform module to deploy Airflow, there are some additional Python packages that you'll need to install in the Python environment of your Airflow server:

• pydantic~=2.7.1: The Airflow DAG files that ZenML creates for you require Pydantic to parse and validate configuration files.

• apache-airflow-providers-docker or apache-airflow-providers-cncf-kubernetes, depending on which Airflow operator you'll be using to run your pipeline steps. Check out this section for more information on supported operators.

How to use it

To use the Airflow orchestrator, we need:

• The ZenML airflow integration installed. If you haven't done so, run

  zenml integration install airflow

• Docker installed and running.
• The orchestrator registered and part of our active stack:

zenml orchestrator register <ORCHESTRATOR_NAME> \
    --flavor=airflow \
    --local=True  # set this to `False` if using a remote Airflow deployment

# Register and activate a stack with the new orchestrator
zenml stack register <STACK_NAME> -o <ORCHESTRATOR_NAME> ... --set

Local

Due to dependency conflicts, we need to install the Python packages to start a local Airflow server in a separate Python environment.

# Create a fresh virtual environment in which we install the Airflow server dependencies
python -m venv airflow_server_environment
source airflow_server_environment/bin/activate

# Install the Airflow server dependencies
pip install "apache-airflow==2.4.0" "apache-airflow-providers-docker<3.8.0" "pydantic~=2.7.1"

Before starting the local Airflow server, we can set a few environment variables to configure it:

• AIRFLOW_HOME: This variable defines the location where the Airflow server stores its database and configuration files. The default value is ~/airflow.

• AIRFLOW__CORE__DAGS_FOLDER: This variable defines the location where the Airflow server looks for DAG files. The default value is <AIRFLOW_HOME>/dags.

• AIRFLOW__SCHEDULER__DAG_DIR_LIST_INTERVAL: This variable controls how often the Airflow scheduler checks for new or updated DAGs. By default, the scheduler will check for new DAGs every 30 seconds. This variable can be used to increase or decrease the frequency of the checks, depending on the specific needs of your pipeline.

When running this on MacOS, you might need to set the no_proxy environment variable to prevent crashes due to a bug in Airflow (see this page for more information):

export no_proxy=*

We can now start the local Airflow server by running the following command:

# Switch to the Python environment that has Airflow installed before running this command
airflow standalone

This command will start up an Airflow server on your local machine. During the startup, it will print a username and password which you can use to log in to the Airflow UI here.

We can now switch back the Python environment in which ZenML is installed and run a pipeline:

# Switch to the Python environment that has ZenML installed before running this command
python file_that_runs_a_zenml_pipeline.py

This call will produce a .zip file containing a representation of your ZenML pipeline for Airflow. The location of this .zip file will be in the logs of the command above. We now need to copy this file to the Airflow DAGs directory, from where the local Airflow server will load it and run your pipeline (It might take a few seconds until the pipeline shows up in the Airflow UI). To figure out the DAGs directory, we can run airflow config get-value core DAGS_FOLDER while having our Python environment with the Airflow installation active.

To make this process easier, we can configure our ZenML Airflow orchestrator to automatically copy the .zip file to this directory for us. To do so, run the following command:

# Switch to the Python environment that has ZenML installed before running this command
zenml orchestrator update --dag_output_dir=<AIRFLOW_DAG_DIRECTORY>

Now that we've set this up, running a pipeline in Airflow is as simple as just running the Python file:

# Switch to the Python environment that has ZenML installed before running this command
python file_that_runs_a_zenml_pipeline.py

Remote

When using the Airflow orchestrator with a remote deployment, you'll additionally need:

• A remote ZenML server deployed to the cloud. See the deployment guide for more information.
• A deployed Airflow server. See the deployment section for more information.
• A remote artifact store as part of your stack.
• A remote container registry as part of your stack.

In the remote case, the Airflow orchestrator works differently than other ZenML orchestrators. Executing a python file which runs a pipeline by calling pipeline.run() will not actually run the pipeline, but instead will create a .zip file containing an Airflow representation of your ZenML pipeline. In one additional step, you need to make sure this zip file ends up in the DAGs directory of your Airflow deployment.

ZenML will build a Docker image called <CONTAINER_REGISTRY_URI>/zenml:<PIPELINE_NAME> which includes your code and use it to run your pipeline steps in Airflow. Check out this page if you want to learn more about how ZenML builds these images and how you can customize them.

Scheduling

You can schedule pipeline runs on Airflow similarly to other orchestrators. However, note thatAirflow schedules always need to be set in the past, e.g.,:

from datetime import datetime, timedelta

from zenml.pipelines import Schedule

scheduled_pipeline = fashion_mnist_pipeline.with_options(
    schedule=Schedule(
        start_time=datetime.now() - timedelta(hours=1),  # start in the past
        end_time=datetime.now() + timedelta(hours=1),
        interval_second=timedelta(minutes=15),  # run every 15 minutes
        catchup=False,
    )
)
scheduled_pipeline()

Airflow UI

Airflow comes with its own UI that you can use to find further details about your pipeline runs, such as the logs of your steps. For local Airflow, you can find the Airflow UI at http://localhost:8080 by default.

If you cannot see the Airflow UI credentials in the console, you can find the password in <AIRFLOW_HOME>/standalone_admin_password.txt. AIRFLOW_HOME will usually be ~/airflow unless you've manually configured it with the AIRFLOW_HOME environment variable. You can always run airflow info to figure out the directory for the active environment.

The username will always be admin.

Additional configuration

For additional configuration of the Airflow orchestrator, you can pass AirflowOrchestratorSettings when defining or running your pipeline. Check out the SDK docs for a full list of available attributes and this docs page for more information on how to specify settings.

Enabling CUDA for GPU-backed hardware

Note that if you wish to use this orchestrator to run steps on a GPU, you will need to follow the instructions on this page to ensure that it works. It requires adding some extra settings customization and is essential to enable CUDA for the GPU to give its full acceleration.

Using different Airflow operators

Airflow operators specify how a step in your pipeline gets executed. As ZenML relies on Docker images to run pipeline steps, only operators that support executing a Docker image work in combination with ZenML. Airflow comes with two operators that support this:

• the DockerOperator runs the Docker images for executing your pipeline steps on the same machine that your Airflow server is running on. For this to work, the server environment needs to have the apache-airflow-providers-docker package installed.

• the KubernetesPodOperator runs the Docker image on a pod in the Kubernetes cluster that the Airflow server is deployed to. For this to work, the server environment needs to have the apache-airflow-providers-cncf-kubernetes package installed.

You can specify which operator to use and additional arguments to it as follows:

from zenml import pipeline, step
from zenml.integrations.airflow.flavors.airflow_orchestrator_flavor import AirflowOrchestratorSettings

airflow_settings = AirflowOrchestratorSettings(
    operator="docker",  # or "kubernetes_pod"
    # Dictionary of arguments to pass to the operator __init__ method
    operator_args={}
)

# Using the operator for a single step
@step(settings={"orchestrator": airflow_settings})
def my_step(...):


# Using the operator for all steps in your pipeline
@pipeline(settings={"orchestrator": airflow_settings})
def my_pipeline(...):

If you're using apache-airflow-providers-cncf-kubernetes>=10.0.0, the import of the Kubernetes pod operator changed, and you'll need to specify the operator like this:

airflow_settings = AirflowOrchestratorSettings(
    operator="airflow.providers.cncf.kubernetes.operators.pod.KubernetesPodOperator"
)

Custom operators

If you want to use any other operator to run your steps, you can specify the operator in your AirflowSettings as a path to the python operator class:

from zenml.integrations.airflow.flavors.airflow_orchestrator_flavor import AirflowOrchestratorSettings

airflow_settings = AirflowOrchestratorSettings(
    # This could also be a reference to one of your custom classes.
    # e.g. `my_module.MyCustomOperatorClass` as long as the class
    # is importable in your Airflow server environment
    operator="airflow.providers.docker.operators.docker.DockerOperator",
    # Dictionary of arguments to pass to the operator __init__ method
    operator_args={}
)

Custom DAG generator file

To run a pipeline in Airflow, ZenML creates a Zip archive that contains two files:

• A JSON configuration file that the orchestrator creates. This file contains all the information required to create the Airflow DAG to run the pipeline.

• A Python file that reads this configuration file and actually creates the Airflow DAG. We call this file the DAG generator and you can find the implementation here .

If you need more control over how the Airflow DAG is generated, you can provide a custom DAG generator file using the setting custom_dag_generator. This setting will need to reference a Python module that can be imported into your active Python environment. It will additionally need to contain the same classes (DagConfiguration and TaskConfiguration) and constants (ENV_ZENML_AIRFLOW_RUN_ID, ENV_ZENML_LOCAL_STORES_PATH and CONFIG_FILENAME) as the original module . For this reason, we suggest starting by copying the original and modifying it according to your needs.

Check out our docs on how to apply settings to your pipelines here.

For more information and a full list of configurable attributes of the Airflow orchestrator, check out the SDK Docs .