AirFly: Auto Generate Airflow's dag.py On The Fly

Pipeline management is essential for data operation in company, many engineering teams rely on tools like Airflow to help them organize workflows, such as ETL, data analytic jobs or machine learning projects.

Airflow provides rich extensibility to let developers arrange workloads into a sequence of "operators", then they declare the task dependencies within a DAG context while writing the dag.py file.

As workflow grows progressively, the increasing complexity of task relations prones to messing up the dag structure, leads to decrease of code maintainability, especially in collaborative scenarios.

airfly tries to mitigate such pain-points and brings automation to this development life cycle, it assumes all tasks are managed in certain python module, developers specify the dependencies while defining the task objects. During deployment, airfly can resolve the dependency tree and automatically build the dag.py for you.

Install

Download airfly from PyPI

$ pip install airfly

$ airfly --help
Usage: airfly [OPTIONS]

Options:
  --version                   Show version and exit.
  -n, --name TEXT             Assign to DAG id.
  -m, --modname TEXT          Name of the module to search tasks for building
                              the task dependency tree and using it to
                              generate the airflow DAG file.
  -p, --path TEXT             Insert into "sys.path" to include certain
                              modules, multi-value is allowed.
  -e, --exclude-pattern TEXT  Exclude the tasks from the dependency tree if
                              their __qualname__ get matched with this regex
                              pattern.
  -i, --includes TEXT         Paths of python files, the code within will be
                              included in the output DAG file, multi-value is
                              allowed.
  -d, --dag-params TEXT       Parameters to construct DAG object, defined by a
                              dictionary in a python file. Pass this option
                              with <python-file>:<variable> form, the
                              <variable> should be the dictionary which will
                              be passed to DAG as keyword arguments.
  --help                      Show this message and exit.

Usage

airfly expects the implementations are populated in a Python module(or package), the task dependencies are declared by assigning upstreams and downstreams attributes to each object. The task objects are actually wrappers for Airflow operators, when airfly walks through the entire module, all tasks are discovered and collected, the dependency tree and the DAG context are automatically built, with some ast helpers, airfly can wrap all these information, convert them into python code, and finally save them to dag.py.

Wrap Airflow operator with AirflowTask

In order to do codegen, collect the operator's metadata into a AirflowTask subclass as following(see demo):

# in demo.py
from airfly.model.airflow import AirflowTask


class print_date(AirflowTask):
    operator_class = "BashOperator" 
    params = dict(bash_command="date")

• operator_class specifies the class of the Airflow operator.
• The class name(print_date) will be mapped to task_id to the applied operator after code generation.
• params will be passed to operator as keyword argument.

Declare task dependency

Use upstreams and downstreams to specify task dependencies.

# in demo.py

from textwrap import dedent


templated_command = dedent(
    """
{% for i in range(5) %}
    echo "{{ ds }}"
    echo "{{ macros.ds_add(ds, 7)}}"
    echo "{{ params.my_param }}"
{% endfor %}
"""
)

class templated(AirflowTask):
    operator_class = "BashOperator"
    params = dict(depends_on_past=False,
                  bash_command=templated_command,
                  params={"my_param": "Parameter I passed in"})


class sleep(AirflowTask):
    operator_class = "BashOperator"
    params = dict(depends_on_past=False, 
                  bash_command="sleep 5",
                  retries=3)

    upstreams = print_date
    downstreams = (templated,)

Generate the dag.py file

With commandline interface:

$ airfly --name demo_dag --modname demo > dag.py

The outputs in dag.py:

# This file is auto-generated by airfly 0.6.0
from airflow.models import DAG
from airflow.operators.bash import BashOperator

with DAG("demo_dag") as dag:
    demo_print_date = BashOperator(task_id="demo.print_date", bash_command="date")
    demo_sleep = BashOperator(
        task_id="demo.sleep", depends_on_past=False, bash_command="sleep 5", retries=3
    )
    demo_templated = BashOperator(
        task_id="demo.templated",
        depends_on_past=False,
        bash_command="""
{% for i in range(5) %}
    echo "{{ ds }}"
    echo "{{ macros.ds_add(ds, 7)}}"
    echo "{{ params.my_param }}"
{% endfor %}
""",
        params={"my_param": "Parameter I passed in"},
    )
    demo_print_date >> demo_sleep
    demo_sleep >> demo_templated

Inject parameters to DAG

If any additional arguments are needed, write and manage those configurations in a python file(see demo), airfly can pass them to DAG during codegen.

# in params.py

from datetime import timedelta

from airflow.utils.dates import days_ago

default_args = {
    "owner": "airflow",
    "depends_on_past": False,
    "email": ["airflow@example.com"],
    "email_on_failure": False,
    "email_on_retry": False,
    "retries": 1,
    "retry_delay": timedelta(minutes=5),
    # 'queue': 'bash_queue',
    # 'pool': 'backfill',
    # 'priority_weight': 10,
    # 'end_date': datetime(2016, 1, 1),
    # 'wait_for_downstream': False,
    # 'dag': dag,
    # 'sla': timedelta(hours=2),
    # 'execution_timeout': timedelta(seconds=300),
    # 'on_failure_callback': some_function,
    # 'on_success_callback': some_other_function,
    # 'on_retry_callback': another_function,
    # 'sla_miss_callback': yet_another_function,
    # 'trigger_rule': 'all_success'
}

dag_kwargs = dict(
    default_args=default_args,
    description="A simple tutorial DAG",
    schedule_interval=timedelta(days=1),
    start_date=days_ago(2),
    tags=["example"],
)

Inject those arguments with --dag-params option:

$ airfly --name demo_dag --modname demo --dag-params params.py:dag_kwargs > dag.py

The outputs in dag.py:

# This file is auto-generated by airfly 0.6.0
from datetime import timedelta

from airflow.models import DAG
from airflow.operators.bash import BashOperator
from airflow.utils.dates import days_ago

# >>>>>>>>>> Include from 'params.py'
default_args = {
    "owner": "airflow",
    "depends_on_past": False,
    "email": ["airflow@example.com"],
    "email_on_failure": False,
    "email_on_retry": False,
    "retries": 1,
    "retry_delay": timedelta(minutes=5),
}
dag_kwargs = dict(
    default_args=default_args,
    description="A simple tutorial DAG",
    schedule_interval=timedelta(days=1),
    start_date=days_ago(2),
    tags=["example"],
)
# <<<<<<<<<< End of code insertion
with DAG("demo_dag", **dag_kwargs) as dag:
    demo_print_date = BashOperator(task_id="demo.print_date", bash_command="date")
    demo_sleep = BashOperator(
        task_id="demo.sleep", depends_on_past=False, bash_command="sleep 5", retries=3
    )
    demo_templated = BashOperator(
        task_id="demo.templated",
        depends_on_past=False,
        bash_command="""
{% for i in range(5) %}
    echo "{{ ds }}"
    echo "{{ macros.ds_add(ds, 7)}}"
    echo "{{ params.my_param }}"
{% endfor %}
""",
        params={"my_param": "Parameter I passed in"},
    )
    demo_print_date >> demo_sleep
    demo_sleep >> demo_templated

airfly wraps required information including variables and imports into output python script, and pass the specified value to DAG object.

Exclude tasks from codegen

By passing --exclude-pattern to match any unwanted objects with their __qualname__. then filter them out.

$ airfly --name demo_dag --modname demo --exclude-pattern templated > dag.py

The outputs in dag.py:

# This file is auto-generated by airfly 0.6.0
from airflow.models import DAG
from airflow.operators.bash import BashOperator

with DAG("demo_dag") as dag:
    demo_print_date = BashOperator(task_id="demo.print_date", bash_command="date")
    demo_sleep = BashOperator(
        task_id="demo.sleep", depends_on_past=False, bash_command="sleep 5", retries=3
    )
    demo_print_date >> demo_sleep

The templated task is gone.

Examples

Please visit examples to explore more use cases.