Airflow Guides

View all Guides


The astro library is an open source Python package maintained by Astronomer that provides tools to improve the DAG authoring experience for Airflow users. The available decorators and functions allow you to write DAGs based on how you want your data to move by simplifying the data transformation process between different environments.

In this guide, we’ll demonstrate how you can use astro functions for ETL use cases. The resulting DAGs will be easier to write and read, and require less code.

Astro ETL Functionality

The astro library makes implementing ETL use cases easier by allowing you to seamlessly transition between Python and SQL for each step in your process. Details like creating dataframes, storing intermediate results, passing context and data between tasks, and creating Airflow task dependencies are all managed automatically. This means that you can focus solely on writing execution logic in whichever language you need without having to worry about Airflow orchestration logic.

More specifically, astro has the following functions that are helpful when implementing an ETL framework (for a full list of functions, check out the Readme):

  • load_file: If the data you’re starting with is in CSV or parquet files (stored locally or on S3 or GCS), you can use this function to load it into your database.
  • transform: This function allows you to transform your data with a SQL query. It uses a SELECT statement that you define to automatically store your results in a new table. By default, the output_table is placed in a tmp_astro schema and is given a unique name each time the DAG runs, but you can overwrite this behavior by defining a specific output_table in your function. You can then pass the results of the transform downstream to the next task as if it were a native Python object.
  • dataframe: Similar to transform for SQL, the dataframe function allows you to implement a transformation on your data using Python. You can easily store the results of the dataframe function in your database by specifying an output_table, which is useful if you want to switch back to SQL in the next step or load your final results to your database.

In the next section, we’ll show a practical example implementing these functions.

Example ETL Implementation

To show astro for ETL in action, we’ll start with a pretty common use case: managing billing data. In this first scenario, we need to extract customer subscription data by joining data from a CSV on S3 with the results of a query to our Snowflake database. Then, we need to perform some transformations on the data before loading it into a results table. First we'll show how to implement this without astro, and then we'll show how astro can make it easier.

The DAG Before Astro

Here is our billing subscription ETL DAG implemented with OSS Airflow operators and decorators, as well as the TaskFlow API:

from datetime import datetime
import pandas as pd

from airflow.decorators import dag, task
from airflow.providers.snowflake.hooks.snowflake import SnowflakeHook
from airflow.providers.snowflake.transfers.s3_to_snowflake import S3ToSnowflakeOperator
from import S3Hook

S3_BUCKET = 'bucket_name'
S3_FILE_PATH = '</path/to/file/'
SNOWFLAKE_CONN_ID = 'snowflake'
SNOWFLAKE_SCHEMA = 'schema_name'
SNOWFLAKE_STAGE = 'stage_name'
SNOWFLAKE_WAREHOUSE = 'warehouse_name'
SNOWFLAKE_DATABASE = 'database_name'
SNOWFLAKE_ROLE = 'role_name'

def extract_data():
    # Join data from two tables and save to dataframe to process
    query = ''''
    SELECT * FROM billing_data
    LEFT JOIN subscription_data
    ON customer_id=customer_id
    # Make connection to Snowflake and execute query
    hook = SnowflakeHook(snowflake_conn_id=SNOWFLAKE_CONN_ID)
    conn = hook.get_conn()
    cur = conn.cursor()

    results = cur.fetchall()
    column_names = list(map(lambda t: t[0], cur.description))

    df = pd.DataFrame(results)
    df.columns = column_names

    return df.to_json()

def transform_data(xcom: str) -> str:
    # Transform data by pivoting
    df = pd.read_json(xcom)

    transformed_df = df.pivot_table(index='DATE', 

    transformed_str = transformed_df.to_string()

    # Save results to S3 so they can be loaded back to Snowflake
    s3_hook = S3Hook(aws_conn_id="s3_conn")
    s3_hook.load_string(transformed_str, 'transformed_file_name.csv', bucket_name=S3_BUCKET, replace=True)

@dag(start_date=datetime(2021, 12, 1), schedule_interval='@daily', catchup=False)

def classic_billing_dag():

    load_subscription_data = S3ToSnowflakeOperator(
        s3_keys=[S3_FILE_PATH + '/subscription_data.csv'],
        file_format="(type = 'CSV',field_delimiter = ',')",

    load_transformed_data = S3ToSnowflakeOperator(
        s3_keys=[S3_FILE_PATH + '/trasnformed_file_name.csv'],
        file_format="(type = 'CSV',field_delimiter = ',')",

    extracted_data = extract_data()
    transformed_data = transform_data(extracted_data)
    load_subscription_data >> extracted_data >> transformed_data >> load_transformed_data

classic_billing_dag = classic_billing_dag()

Classic Graph

While we achieved our ETL goal with the DAG above, there are a couple of limitations that made this implementation more complicated:

  • Since there is no way to pass results from SnowflakeOperator query to the next task, we had to write our query in a _DecoratedPythonOperator function using the SnowflakeHook and explicitly do the conversion from SQL to a dataframe ourselves.
  • Some of our transformations are better suited to SQL, and others are better suited to Python, but transitioning between the two requires extra boilerplate code to explicitly make those conversions.
  • While the TaskFlow API makes it easier to pass data between tasks here, it is storing the resulting dataframes as XComs by default. This means that we need to worry about the size of our data. We could implement a custom XCom backend, but that would be extra lift.
  • Loading data back to Snowflake after the transformation is complete requires writing extra code to store an intermediate CSV in S3.

The DAG With Astro

Next, we’ll show how to rewrite the DAG using astro to alleviate the challenges listed above.

from airflow.decorators import dag
from astro.sql import transform, append, load_file
from astro.sql.table import Table
from astro import dataframe

from datetime import datetime
import pandas as pd

SNOWFLAKE_CONN_ID = "snowflake"
S3_FILE_PATH = '</path/to/file/'

# Start by selecting data from two source tables in Snowflake
def extract_data(subscriptions: Table, customer_data: Table):
    return """
    SELECT *
    FROM {subscriptions}
    LEFT JOIN {customer_data}
           ON customer_id=customer_id

# Switch to Pandas for pivoting transformation
def transform_data(df: pd.DataFrame):
    transformed_df = df.pivot_table(index='DATE', 

    return transformed_df

main_table = Table("billing_reporting", schema="SANDBOX_KENTEND", )

@dag(start_date=datetime(2021, 12, 1), schedule_interval='@daily', catchup=False)

def astro_billing_dag():
    # Load subscripton data
    subscription_data = load_file(
        path=S3_FILE_PATH + '/subscription_data.csv',
        output_table=Table(table_name="subscription_data", conn_id=SNOWFLAKE_CONN_ID),
    # Define task dependencies
    extracted_data = extract_data(
        customer_data=Table('customer_data', schema='SANDBOX_KENTEND')

    transformed_data = transform_data(
    # Append transformed data to billing table
    # Dependency is inferred by passing the previous `transformed_data` task to `append_table` param
        columns=["DATE", "CUSTOMER_ID", "AMOUNT"],
        main_table=Table("billing_reporting", schema="SANDBOX_KENTEND"),

astro_billing_dag = astro_billing_dag()

Astro Graph

The key differences in this implementation are:

  • The load_file and append functions take care of loading our raw data from S3 and transforming data to our reporting table respectively. We didn't have to write any extra code to get the data into Snowflake.
  • Using the transform function, we easily executed SQL to combine our data from multiple tables. The results are automatically stored in a table in Snowflake. We didn't have to use the SnowflakeHook or write any of the code to execute the query.
  • We seamlessly transitioned to a transformation in Python with the df function without needing to explicitly convert the results of our previous task to a Pandas dataframe. We then wrote the output of our transformation to our aggregated bills table in Snowflake using the output_table parameter, so we didn't have to worry about storing the data in XCom.
  • We only had to define our connection in the first load_file task. All downstream task that inherit from a task with a connection defined will use it.

Overall, our DAG with astro is shorter, simpler to implement, and easier to read. This allows us to implement even more complicated use cases easily while focusing on the movement of our data.