Making DAG construction easier


Keywords
airflow, data-etl, data-pipeline
License
MIT
Install
pip install gusty==0.22.0

Documentation

logo

Versions PyPi build coverage Code Style


gusty allows you to control your Airflow DAGs, Task Groups, and Tasks with greater ease. gusty manages collections of tasks, represented as any number of YAML, Python, SQL, Jupyter Notebook, or R Markdown files. A directory of task files is instantly rendered into a DAG by passing a file path to gusty's create_dag function.

gusty also manages dependencies (within one DAG) and external dependencies (dependencies on tasks in other DAGs) for each task file you define. All you have to do is provide a list of dependencies or external_dependencies inside of a task file, and gusty will automatically set each task's dependencies and create external task sensors for any external dependencies listed.

gusty works with both Airflow 1.x and Airflow 2.x, and has even more features, all of which aim to make the creation, management, and iteration of DAGs more fluid, so that you can intuitively design your DAG and build your tasks.

Documentation

The official documentation for gusty is hosted here: https://pipeline-tools.github.io/gusty-docs/

What's in gusty?

Five Ways to Make Tasks

gusty will turn every file in a DAG directory into a task. By default gusty supports five different file types, which offer convenient ways to specify an operator and operator parameters for task creation.

File Type How It Works
.yml Declare an operator and pass in any operator parameters using YAML
.py Simply write Python code and by default gusty will execute your file using a PythonOperator. Other options available
.sql Declare an operator in a YAML header, then write SQL in the main .sql file. The SQL automatically gets sent to the operator
.ipynb Put a YAML block at the top of your notebook and specify an operator that renders your Jupyter Notebook
.Rmd Use the YAML block at the top of your notebook and specify an operator that renders your R Markdown Document

Here is quick example of a YAML task file, which might be called something like hello_world.yml:

operator: airflow.operators.bash.BashOperator
bash_command: echo hello world

The resulting task would be a BashOperator with the task id hello_world.

Here is the same approach using a Python file instead, named hello_world.py, which gusty will automatically turn into a PythonOperator by default:

phrase = "hello world"
print(phrase)

Lastly, here's a slightly different .sql example:

---
operator: airflow.providers.sqlite.operators.sqlite.SqliteOperator
---

SELECT
  column_1,
  column_2
FROM your_table

Easy Dependencies

Declarative Dependencies

Every task file type supports dependencies and external_dependencies parameters, which gusty will use to automatically assign dependencies between tasks and create external task sensors for any external dependencies listed for a given task.

For .yml, .ipynb, and .Rmd task file types, dependencies and external_dependencies would be defined using YAML syntax:

operator: airflow.operators.bash.BashOperator
bash_command: echo hello world
dependencies:
  - same_dag_task
external_dependencies:
  - another_dag: another_task
  - a_whole_dag: all

For external dependencies, the keyword all can be used when the task should wait on an entire external DAG to run successfully.

For a .py task file type, we can define these dependencies with some raw markdown at the top of the file:

# ---
# dependencies:
#   - same_dag_task
# external_dependencies:
#   - another_dag: another_task
#   - a_whole_dag: all
# python_callable: say_hello
# ---

def say_hello():
  phrase = "hello world"
  print(phrase)

You will also note that we wrapped our previous Python code in a function called say_hello, and passed this function's name to the python_callable argument. By default, with no operator and no python_callable specified, gusty will pass a simple function that runs your .py file to the PythonOperator. If you pass an explicit python_callable by name, gusty will search your .py for that function and pass that function to the PythonOperator instead.

.py files are capable of accepting an operator parameter in the raw markdown, just like any other task file type, which means you can use any other relevant operators (e.g. the PythonVirtualenvOperator) to execute your Python code as needed.

Dynamic Dependencies

gusty can also detect and generate dependencies through a task object's dependencies attribute. This means you can also dynamically set dependencies. One popular example of this option would be if your operator runs SQL, you can parse that SQL for table names, and attach a list of those table names to the operator's dependencies attribute. If those table names listed in the dependencies attribute are also task ids in the DAG, gusty will be able to automatically set these dependencies for you!

DAG and TaskGroup Control

Both DAG and TaskGroup objects are created automatically simply by being directories and subfolders, respectively. The directory path you provide to gusty's create_dag function will become your DAG (and DAG name), and any subfolder in that DAG by default will be turned into a TaskGroup.

gusty offers a few compatible methods for configuring DAGs and Task Groups that we'll cover below.

Metadata

A special file name in any directory or subfolder is METADATA.yml, which gusty will use to determine how to configure that DAG or TaskGroup object.

Here is an example of a METADATA.yml file you might place in a DAG directory:

description: "An example of a DAG created using METADATA.yml"
schedule_interval: "1 0 * * *"
default_args:
    owner: airflow
    depends_on_past: False
    start_date: !days_ago 1
    email: airflow@example.com
    email_on_failure: False
    email_on_retry: False
    retries: 1
    retry_delay: !timedelta 'minutes: 5'

And here is an example of a METADATA.yml file you might place in a TaskGroup subfolder:

tooltip: "This is a task group tooltip"
prefix_group_id: True
dependencies:
  - hello_world

As seen in the above example, gusty will also accept dependencies and external_dependencies in a TaskGroup's METADATA.yml. This means gusty can wire up your TaskGroup dependencies as well!

Note that gusty disables the TaskGroup prefix_group_id argument by default, as it's one of gusty's few opinions that tasks should explicitly named unless you say otherwise. gusty also offers a suffix_group_id argument for Task Groups!

create_dag

While METADATA.yml will always be the primary source of truth for a DAG or TaskGroup's configuration, gusty's create_dag function also accepts any parameters that can be passed to Airflow's DAG class, as well as a dictionary of task_group_defaults to set default behavior for any TaskGroup created by gusty.

Here's an example using create_dag, where instead of metadata we use create_dag arguments:

from datetime import timedelta
from airflow.utils.dates import days_ago
from gusty import create_dag

dag = create_dag(
  '/usr/local/airflow/dags/hello_world',
  description="A DAG created without any metadata",
  schedule_interval="1 0 * * *",
  default_args={
      "owner": "airflow",
      "depends_on_past": False,
      "start_date": days_ago(1),
      "email": "airflow@example.com",
      "email_on_failure": False,
      "email_on_retry": False,
      "retries": 1,
      "retry_delay": timedelta(minutes=5),
  },
  task_group_defaults={
      "tooltip": "This is a task group tooltip",
      "prefix_group_id": True
  }
)

You might notice that task_group_defaults does not include dependencies. For Task Groups, dependencies must be set using TaskGroup-specific metadata.

Default arguments in create_dag and a DAG or TaskGroup's METADATA.yml can be mixed and matched. METADATA.yml will always override defaults set in create_dag.

create_dags

If you have multiple gusty DAGs located inside of a single directory, you can conveniently use the create_dags (plural) function.

create_dags works just like create_dag, with two exceptions:

  1. The first argument to create_dags is the path to a directory with many gusty DAGs.

  2. The second argument to create_dags is globals(). globals() is essentially the namespace to which your DAGs are assigned.

Let's adjust the above create_dag example to use create_dags instead:

from datetime import timedelta
from airflow.utils.dates import days_ago
from gusty import create_dags

create_dags(
  '/usr/local/airflow/my_gusty_dags',
  globals(),
  description="A default description for my DAGs.",
  schedule_interval="1 0 * * *",
  default_args={
      "owner": "airflow",
      "depends_on_past": False,
      "start_date": days_ago(1),
      "email": "airflow@example.com",
      "email_on_failure": False,
      "email_on_retry": False,
      "retries": 1,
      "retry_delay": timedelta(minutes=5),
  },
  task_group_defaults={
      "tooltip": "This is a task group tooltip",
      "prefix_group_id": True
  }
)

The above will create many gusty DAGs located in the /usr/local/airflow/my_gusty_dags directory.

DAG-level Features

gusty features additional helpful arguments at the DAG-level to help you design your DAGs with ease:

  • root_tasks - A list of task ids which should represent the roots of a DAG. For example, an HTTP sensor might have to succeed before any downstream tasks in the DAG run.
  • leaf_tasks - A list of task ids which should represent the leaves of a DAG. For example, at the end of the DAG run, you might save a report to S3.
  • external_dependencies - You can also set external dependencies at the DAG level! Making your DAG wait on other DAGs works just like in the external dependencies examples above.
  • ignore_subfolders - If you don't want subfolders to generate Task Groups, set this to True.
  • latest_only - On by default, installs a LatestOnlyOperator at the absolute root of the DAG, skipping all tasks in the DAG if the DAG run is not the current run. You can read more about the LatestOnlyOperator in Airflow's documentation.

Any of these arguments can be placed in create_dag or METADATA.yml!

Local Operator Support

While you can store your local operators in Airflow's plugins directory and reference an operator's plugins path accordingly, gusty also allows you to alternatively keep local operators inside of an operators folder located inside of your AIRFLOW_HOME.

In order for gusty to support your operators as expected, your operator name must be CamelCase and the file in which the operator lives must be snake_case.

For example, if we wanted use a HelloOperator, this operator would need to be stored in a file called hello_operator.py in inside of the operators folder located inside of your AIRFLOW_HOME.

Any fields from your operator's __init__ method will be passed from gusty to your operator. So if your HelloOperator had a name field, you could call this operator with a YAML task file that looks something like this:

operator: local.HelloOperator
name: World

The local. syntax is what gusty uses to know to look in your local operators folder for the operator.

Multi-Task Generation

Sometimes task definitions can be repetitive. To account for this, gusty allows for a multi_task_spec block in any frontmatter. This allows you to generate multiple similar tasks with a single task definition file! For example, let's say you wanted to create two bash tasks, each containing a different bash_command. You can define these two tasks in a single task definition file like so:

operator: airflow.operators.bash.BashOperator
multi_task_spec:
  bash_task_1:
    bash_command: echo first_task
  bash_task_2:
    bash_command: echo second_task

gusty will convert the above into two task instances, bash_task_1 and bash_task_2, each with a unique bash_command.

Additionally, for the special case of python_callable in a .py file, you can specify python_callable_partials:

# ---
# python_callable: main
# python_callable_partials:
#   python_task_1:
#       my_kwarg: a
#   python_task_2:
#       my_kwarg: b
# ---


def main(my_kwarg):
    return my_kwarg

gusty will convert the above in two task instances, python_task_1 and python_task_2. python_task_1 will return "a" and python_task_2 will return "b".

multi_task_spec and python_callable_partials are non-exclusive, so you can mix and match configuration as needed.

One Approach, Not the Only Approach

One good thing about gusty is that if you choose to use this package, gusty doesn't have to be the only way that you create DAGs. You can use gusty's create_dag to generate DAGs out of directories where applicable, and then implement more traditional methods of creating Airflow DAGs where the tried and true methods feel like a better approach.

So feel free to give the gusty approach a try, because you don't have to commit to it everywhere. But when you try it, don't be surprised if you start using it everywhere!

Containerized Demo

As an additional resource, you can check out a containerized demo of gusty and Airflow over at the gusty-demo repo, which illustrates how gusty and a few custom operators can make SQL queries, Jupyter notebooks, and RMarkdown documents all work together in the same data pipeline.

Development

The below assumes you have Docker installed.

Starting for the First Time

First, git clone this repository.

Then:

export GUSTY_DEV_HOME="~/path/to/this/project"
cd $GUSTY_DEV_HOME
make build-image
make run-image

The above will build the development image under the name gusty-testing and run a container called gusty-testing.

From here, you can:

  • make exec - Exec into a terminal in the running container.
  • make test - Runs pytest in a temporary container.
  • make coverage - Runs pytest and generates a coverage report.
  • make browse-coverage - Opens up the coverage report in your browser.
  • make stop-container - Stop the running container.
  • make start-container - Start a stopped container.

Rebuilding the Image

make stop-container # if you have a running container
make remove-container # if you have a stopped container
make build-image
make run-image