Apache Airflow is a platform to programmatically author, schedule and monitor workflows. It is a commonly used framework for building model training pipelines in ML projects. BentoML provides a flexible set of APIs for integrating natively with Apache Airflow. Users can use Airflow to schedule their model training pipelines and use BentoML to keep tracked of trained model artifacts and optionally deploy them to production in an automated fashion.

This is especially userful for teams that can benefit from retraining models often with newly arrived data, and want to update their production models regularly with confidence.

For more in-depth Airflow tutorials, please visit the Airflow documentation.


A typical Airflow pipeline with a BentoML serving & deployment workflow look like this:

  1. Fetch new data batches from a data source

  2. Split the data in train and test sets

  3. Perform feature extraction on training data set

  4. Train a new model using the training data set

  5. Perform model evaluation and validation

  6. Save the model with BentoML

  7. Push the saved model to BentoCloud (or export model to s3)

  8. Build a new Bento using the newly trained model

  9. Run integration test on the Bento to verify the entire serving pipeline

  10. Push the Bento to BentoCloud (or export bento to s3)

  11. (Optional) Trigger a redeployment via BentoCloud or a custom deployment script

Pro Tips#

Pipeline Dependencies#

The default PythonOperator requires all the dependencies to be installed on the Airflow environment. This can be challenging to manage when the pipeline is running on a remote Airflow deployment and running a mix of different tasks.

To avoid this, we recommend managing dependencies of your ML pipeline with the PythonVirtualenvOperator, which runs your code in a virtual environment. This allows you to define your Bento’s dependencies in a requirements.txt file and use it across training pipeline and the bento build process. For example:

from datetime import datetime, timedelta
from airflow import DAG
from airflow.decorators import task

with DAG(
    description='A simple tutorial DAG with BentoML',
    start_date=datetime(2021, 1, 1),
) as dag:

    def build_bento(**context):
        Perform Bento build in a virtual environment.
        import bentoml
        bento =
                "job_id": context.run_id
                requirements_txt: "./requirements.txt"

    build_bento_task = build_bento()

Artifact Management#

Since Airflow is a distributed system, it is important to save the models and Bentos produced in your Airflow pipeline to a central location that is accessible by all the nodes in the Airflow cluster, and also by the workers in your production deployment environment.

For a simple setup, we recommend using the Import/Export API for models and Bentos. This allows you to export the model files directly to cloud storage, and import them from the same location when needed. E.g:



For a more advanced setup, we recommend using the Model and Bento repositories provided in BentoCloud, which provides additional management features such as filtering, labels, and a web UI for browsing and managing models. E.g:



Python API or CLI#

BentoML provides both Python APIs and CLI commands for most workflow management tasks, such as building Bento, managing Models/Bentos, and deploying to production.

When using the Python APIs, you can organize your code in a Airflow PythonOperator task. And for CLI commands, you can use the BashOperator instead.

Validating new Bento#

It is important to validate the new Bento before deploying it to production. The bentoml.testing module provides a set of utility functions for building behavior tests for your BentoML Service, by launching the API server in a docker container and sending test requests to it.

The BentoML community is also building a standardized way of defining and running test cases for your Bento, that can be easily integrated with your CI/CD pipeline in an Airflow job. See #2967 for the latest progress.

Saving model metadata#

When saving a model with BentoML, you can pass in a dictionary of metadata to be saved together with the model. This can be useful for tracking model evaluation metrics and training context, such as the training dataset timestamp, training code version, or training parameters.

Sample Project#

The following is a sample project created by the BentoML community member Sarah Floris, that demonstrates how to use BentoML with Airflow: