AI / Machine Learning
January 8, 2021

Experimenting with Airflow to Process S3 Files

As machine learning developers, we always need to deal with ETL processing (Extract, Transform, Load) to get data ready for our model. Airflow can help us build ETL pipelines, and visualize the results for each of the tasks in a centralized way.

In this blog post, we look at some experiments using Airflow to process files from S3, while also highlighting the possibilities and limitations of the tool.

What is Airflow?

Airflow is a platform used to programmatically schedule and monitor the workflows of tasks. This workflow is designed as a dependency graph between tasks.

It is composed of a scheduler that sends tasks to be executed following the dependencies and workers that execute the tasks. It also provides a user interface to visualize and monitor running pipelines, view logs, and start workflows manually.

Airflow Executors: Executors are the mechanisms on which task instances get to run. Airflow has different executors, which you can see here. Here are the most common:

  • SequentialExecutor: The default executor where tasks are executed sequentially.
  • LocalExecutor: LocalExecutor runs tasks on the same node as the scheduler. It allows you to locally run multiple jobs in parallel.
  • CeleryExecutor: This is the most mature option and requires Redis or RabbitMQ to queue the tasks.
  • KubernetesExecutor: This executor was introduced in version 1.10.0. With KubernetesExecutor, you can prepare different docker images for your tasks with more flexibility.

FYI - since CeleryExecutor is more mature, experiments have been performed with this executor in the architecture.

Airflow Architecture

So, when using the Celery Executor, these are the componentes of the architecture:

  • Database: This contains DAG's (workflows) status and task instances.
  • Airflow Web Server: A web interface to query the database status, and monitor and execute DAGs.
  • Airflow Scheduler: This sends tasks to the queues and updates information in the database.
  • Message broker: Inserts the task's commands to be run into the queue.
  • Airflow Celery workers: Retrieves commands from the queue, executes them, and updates the database.

So, the Airflow Scheduler uses the Celery Executor to schedule tasks. The Celery Executor enqueues the tasks, and each of the workers takes the queued tasks to be executed.

All of the components are deployed in a Kubernetes cluster. The database can be MySQL or Postgres, and the message broker might be RabbitMQ or Redis.


A DAG (Directed Acyclic Graph) represents a group of tasks, where dependence might exist between them. It is defined as a python script that represents the DAG's structure (tasks and their dependencies) as code.

Each of the tasks is implemented with an operator. Different types of operators exist, and you can create your custom operator if necessary. The most common operators are BashOperator (to execute bash actions), and PythonOperator (to execute python scripts/functions). In this blog post, different examples are provided using some of the operators available.

Running Airflow locally

You can locally run the Kubernetes cluster for airflow locally with Docker Compose. Download this repo.

Docker Compose runs an airflow architecture composed by:
- 1 worker & scheduler
- Flower (jobs' UI)
- Redis (as broker)
- Postgres (database)

So you know, all components are docker containers.

Run with Celery Executor:

[.c-inline-code]docker-compose -f docker-compose-CeleryExecutor.yml up -d

Basic Operations

List DAGs: In the web interface you can list all the loaded DAGs and their state.

You can use the command line to check the configured DAGs:

[.c-inline-code] docker exec -ti docker-airflow_scheduler_1 ls dags/

Run Manually In the list view, activate the DAG with the On/Off button. Then, enter the DAG and press the Trigger button.

See Logs:

To see the logs for a task from the web, click on the task, and press the 'View Log' button.

You can also check the logs for the scheduler and the worker from the console via the following:

  • Scheduler's logs :

[.c-inline-code] docker logs -f docker-airflow_scheduler_1[.c-inline-code]

  • see worker's logs:

[.c-inline-code] docker logs -f docker-airflow_worker_1[.c-inline-code]

Experiments with Operators:


The BashOperator executes a bash command. This example contains three bash tasks, two of which can be executed in parallel. To execute it, activate the tutorial DAG and enter the view for the DAG. You will then see that the first task of the DAG will be scheduled and then queued for completion.


This Operator is used to download files from an S3 bucket, before transforming and then uploading them to another bucket. Therefore, in order to use this operator, we need to configure an S3 connection.

In the web interface, go to Admin->Connections, and set the connection id and type. Add the access key and the secret key as ‘extra’ arguments.

To get the canonical user for S3:

[.c-inline-code]aws s3api list-buckets --query Owner.Ioutput text[.c-inline-code]

from airflow import DAG
from airflow.operators.bash_operator import BashOperator
from airflow.operators.s3_file_transform_operator import S3FileTransformOperator

from datetime import datetime, timedelta

default_args = {
    "owner": "airflow",
    "depends_on_past": False,
    "start_date": datetime(2020, 9, 7),
    "email": [""],
    "email_on_failure": False,
    "email_on_retry": False,
    "retries": 1,
    "retry_delay": timedelta(minutes=5)

with DAG("s3_transformer", default_args=default_args, schedule_interval= '@once') as dag:

    t1 = BashOperator(
        bash_command='echo "hello, it should work" > s3_conn_test.txt'

    transformer = S3FileTransformOperator(
        description='cleans ETL_medical_records',


Change the parameters source_s3_key and dest_s3_key in the script, and then copy the dag to the dags folder.

[.c-inline-code] docker cp  docker-airflow_webserver_1:/usr/local/airflow/dags/

Create the script ''

	import sys
	print("Starting data transformation...")
	print("Completed data transformation!")

Copy it to the container:

[.c-inline-code] docker exec -ti docker-airflow_webserver_1 mkdir /usr/local/airflow/dags/scfipts/ & docker cp  docker-airflow_webserver_1:/usr/local/airflow/dags/scfipts/

To run this task, we will need to install some libraries in the containers and then restart them:

[.c-inline-code] docker exec -ti docker-airflow_worker_1 pip install boto3 boto botocore & docker exec -ti docker-airflow_scheduler_1 pip install boto3 boto botocore & docker exec -ti docker-airflow_webserver_1 pip install boto3 boto botocore

Restart the containers:

[.c-inline-code] docker restart docker-airflow_worker_1 & docker restart docker-airflow_scheduler_1 & docker restart docker-airflow_webserver_1

It's important that the script you set in the S3FileTransformOperator starts with **#!/usr/bin/python3 **in the form of python.

Potential problem: If your script needs specific libraries to be installed (like pandas), they are NOT installed in the worker. So, when it executes, the task then gives you an error. There is not a clean solution for this issue unless you use KubernetesExecutor instead of celery.

If you are having problems, you can create a DAG that contains a S3KeySensor to test the connection.


This connector allows you to make a query to Athena's database. You will need to set the s3_connection in the [.c-inline-code]aws_conn_id[.c-inline-code] parameter. This connection should be defined in the connection configuration.

Update the following script with the correct database and desired query. You need to create a database in AWS Athena to query the S3 files.

from airflow import DAG
from airflow.operators.bash_operator import BashOperator
from airflow.operators.s3_file_transform_operator import S3FileTransformOperator
from airflow.contrib.operators.aws_athena_operator import AWSAthenaOperator

from datetime import datetime, timedelta

default_args = {
    "owner": "airflow",
    "depends_on_past": False,
    "start_date": datetime(2020, 9, 7),
    "email": [""],
    "email_on_failure": False,
    "email_on_retry": False,
    "retries": 1,
    "retry_delay": timedelta(minutes=5)

with DAG("query_s3", default_args=default_args, schedule_interval= '@once') as dag:

    t1 = BashOperator(
        bash_command='echo "Starting AWSAthenaOperator TEST"'

    run_query = AWSAthenaOperator(
        query='select {PARAM} FROM "{DATABASE}"."{TABLE}"',


Copy the DAG to the DAG's directory and execute it from the web interface.


This DAG executes the task into a pod, and you then have the option to kill the pod once it finishes the execution.
Note - if you are seeing the following error:

[.c-inline-code] {} ERROR - Exception when attempting to create Namespaced Pod.

Add in_cluster=True in the DAG to specify that the pod will be running in the same cluster.


This dynamically creates a subDag inside the dag.

from airflow import DAG

from airflow.operators.subdag_operator import SubDagOperator
from airflow.operators.bash_operator import BashOperator
from airflow.hooks.S3_hook import S3Hook
from airflow.operators.dummy_operator import DummyOperator
from airflow.operators.python_operator import PythonOperator

from datetime import datetime, timedelta

default_args = {
    "owner": "airflow",
    "depends_on_past": False,
    "start_date": datetime(2020, 9, 7),
    "email": [""],
    "email_on_failure": False,
    "email_on_retry": False,
    "retries": 1,
    "retry_delay": timedelta(minutes=5)

def hello(file):
    print('Hello!!!! ', file)

def loop_files(parent_dag_name, child_dag_name, args):
    dag_subdag = DAG(
        dag_id='{0}.{1}'.format(parent_dag_name, child_dag_name),
    tasks = []
    for i in range (5):
        tasks = tasks + [PythonOperator(
            task_id='hello_world' + str(i),
            op_kwargs={'file': str(i)},

    return dag_subdag

dag = DAG("subdagtest2", default_args=default_args, schedule_interval= '@once')

start_op = BashOperator(
    bash_command='echo "Starting TEST"',
    dag=dag )

loop_files = SubDagOperator(
    subdag=loop_files('subdagtest2', 'loop_files', default_args),dag=dag

start_op >> loop_files

Once you run it, you will see a web option to enter the SubDag's information and logs:

This example lists the files in an S3 bucket, and for each file, it then creates a SubDAG "hellow_wold_X".

Potential problem: Often times too many tasks are queued, and it is probable that you will need to add more workers.

Lessons Learned

When you don't need specific dependencies, it's better to use BashOperator or PythonOperator

On the other hand, when your tasks need specific dependencies, it's better to use KubernetesOperator.

Also, SubDAGs are useful when you need to repeat a series of tasks for each S3 file. However, you need to ensure that you remember to control the size of the queue.