DAG Authoring Best Practices

DAG Authoring Best Practices

Import Compatibility

Airflow 2.x:

from airflow.decorators import dag, task, task_group, setup, teardown
from airflow.models import Variable
from airflow.hooks.base import BaseHook

Airflow 3.x (Task SDK):

from airflow.sdk import dag, task, task_group, setup, teardown, Variable, Connection

The examples below use Airflow 2 imports for compatibility. On Airflow 3, these still work but are deprecated (AIR31x warnings). For new Airflow 3 projects, prefer airflow.sdk imports.

Table of Contents

• Avoid Top-Level Code
• TaskFlow API
• Credentials Management
• Provider Operators
• Idempotency
• Data Intervals
• Task Groups
• Dynamic Task Mapping
• Large Data / XCom
• Retries and Scaling
• Sensor Modes and Deferrable Operators
• Setup/Teardown
• Data Quality Checks
• Anti-Patterns
• Assets (Airflow 3.x)

Avoid Top-Level Code

DAG files are parsed every ~30 seconds. Code outside tasks runs on every parse.

# WRONG - Runs on every parse (every 30 seconds!)
hook = PostgresHook("conn")
results = hook.get_records("SELECT * FROM table")  # Executes repeatedly!

@dag(...)
def my_dag():
    @task
    def process(data):
        return data
    process(results)

# CORRECT - Only runs when task executes
@dag(...)
def my_dag():
    @task
    def get_data():
        hook = PostgresHook("conn")
        return hook.get_records("SELECT * FROM table")

    @task
    def process(data):
        return data

    process(get_data())

Use TaskFlow API

from airflow.decorators import dag, task  # AF3: from airflow.sdk import dag, task
from datetime import datetime

@dag(
    dag_id='my_pipeline',
    start_date=datetime(2025, 1, 1),
    schedule='@daily',
    catchup=False,
    default_args={'owner': 'data-team', 'retries': 2},
    tags=['etl', 'production'],
)
def my_pipeline():
    @task
    def extract():
        return {"data": [1, 2, 3]}

    @task
    def transform(data: dict):
        return [x * 2 for x in data["data"]]

    @task
    def load(transformed: list):
        print(f"Loaded {len(transformed)} records")

    load(transform(extract()))

my_pipeline()

Never Hard-Code Credentials

# WRONG
conn_string = "postgresql://user:password@host:5432/db"

# CORRECT - Use connections
from airflow.hooks.base import BaseHook  # AF3: from airflow.sdk import Connection
conn = BaseHook.get_connection("my_postgres_conn")

# CORRECT - Use variables
from airflow.models import Variable  # AF3: from airflow.sdk import Variable
api_key = Variable.get("my_api_key")

# CORRECT - Templating
sql = "SELECT * FROM {{ var.value.table_name }}"

Use Provider Operators

from airflow.providers.snowflake.operators.snowflake import SnowflakeOperator
from airflow.providers.google.cloud.operators.bigquery import BigQueryInsertJobOperator
from airflow.providers.common.sql.operators.sql import SQLExecuteQueryOperator

Ensure Idempotency

@task
def load_data(data_interval_start, data_interval_end):
    # Delete before insert
    delete_existing(data_interval_start, data_interval_end)
    insert_new(data_interval_start, data_interval_end)

Use Data Intervals

@task
def process(data_interval_start, data_interval_end):
    print(f"Processing {data_interval_start} to {data_interval_end}")

# In SQL
sql = """
    SELECT * FROM events
    WHERE event_time >= '{{ data_interval_start }}'
      AND event_time < '{{ data_interval_end }}'
"""

Airflow 3 context injection: In Airflow 3 (Task SDK), context variables are automatically injected as function parameters by name. A bare type annotation is valid — no = None default required:

import pendulum

# Airflow 3 — both forms are valid
@task
def process(data_interval_end: pendulum.DateTime):  # No default needed
    ...

@task
def process(data_interval_end: pendulum.DateTime = None):  # Also valid but unnecessary in AF3
    ...

Organize with Task Groups

from airflow.decorators import task_group, task  # AF3: from airflow.sdk import task_group, task

@task_group
def extract_sources():
    @task
    def from_postgres(): ...

    @task
    def from_api(): ...

    return from_postgres(), from_api()

Use Dynamic Task Mapping

Process variable numbers of items in parallel instead of loops:

# WRONG - Sequential, one failure fails all
@task
def process_all():
    for f in ["a.csv", "b.csv", "c.csv"]:
        process(f)

# CORRECT - Parallel execution
@task
def get_files():
    return ["a.csv", "b.csv", "c.csv"]

@task
def process_file(filename): ...

process_file.expand(filename=get_files())

# With constant parameters
process_file.partial(output_dir="/out").expand(filename=get_files())

Handle Large Data (XCom Limits)

For large data, prefer the claim-check pattern: write to external storage (S3, GCS, ADLS) and pass a URI/path reference via XCom.

# WRONG - May exceed XCom limits
@task
def get_data():
    return huge_dataframe.to_dict()  # Could be huge!

# CORRECT - Claim-check pattern: write to storage, return reference
@task
def extract(**context):
    path = f"s3://bucket/{context['ds']}/data.parquet"
    data.to_parquet(path)
    return path  # Small string reference (the "claim check")

@task
def transform(path: str):
    data = pd.read_parquet(path)  # Retrieve data using the reference
    ...

Airflow 3 XCom serialization: Airflow 3's Task SDK natively supports serialization of common Python types including DataFrames. Airflow 2 required a custom XCom backend or manual serialization for non-primitive types.

For automatic offloading, use the Object Storage XCom backend (provider common-io).

AIRFLOW__CORE__XCOM_BACKEND=airflow.providers.common.io.xcom.backend.XComObjectStorageBackend
AIRFLOW__COMMON_IO__XCOM_OBJECTSTORAGE_PATH=s3://conn_id@bucket/xcom
AIRFLOW__COMMON_IO__XCOM_OBJECTSTORAGE_THRESHOLD=1048576
AIRFLOW__COMMON_IO__XCOM_OBJECTSTORAGE_COMPRESSION=gzip

Configure Retries and Scaling

from datetime import timedelta

@dag(
    max_active_runs=1,       # Concurrent DAG runs
    max_active_tasks=10,     # Concurrent tasks per run
    default_args={
        "retries": 3,
        "retry_delay": timedelta(minutes=5),
        "retry_exponential_backoff": True,
    },
)
def my_dag(): ...

# Use pools for resource-constrained operations
@task(pool="db_pool", retries=5)
def query_database(): ...

Environment defaults:

AIRFLOW__CORE__DEFAULT_TASK_RETRIES=2
AIRFLOW__CORE__PARALLELISM=32

Sensor Modes and Deferrable Operators

Prefer deferrable=True when available. Otherwise, use mode='reschedule' for waits longer than a few minutes. Reserve mode='poke' (the default) for sub-minute checks only.

from airflow.providers.amazon.aws.sensors.s3 import S3KeySensor

# WRONG for long waits - poke is the default, so omitting mode= has the same problem
S3KeySensor(
    task_id="wait_for_file",
    bucket_key="data/{{ ds }}/input.csv",
    # mode defaults to "poke" — holds a worker slot the entire time
    poke_interval=300,
    timeout=7200,
)

# CORRECT - frees worker between checks
S3KeySensor(
    task_id="wait_for_file",
    bucket_key="data/{{ ds }}/input.csv",
    mode="reschedule",  # Releases worker between pokes
    poke_interval=300,
    timeout=7200,
)

# BEST - deferrable uses triggerer, most efficient
S3KeySensor(
    task_id="wait_for_file",
    bucket_key="data/{{ ds }}/input.csv",
    deferrable=True,
)

Use Setup/Teardown

from airflow.decorators import dag, task, setup, teardown  # AF3: from airflow.sdk import ...

@setup
def create_temp_table(): ...

@teardown
def drop_temp_table(): ...

@task
def process(): ...

create = create_temp_table()
process_task = process()
cleanup = drop_temp_table()

create >> process_task >> cleanup
cleanup.as_teardown(setups=[create])

Include Data Quality Checks

from airflow.providers.common.sql.operators.sql import (
    SQLColumnCheckOperator,
    SQLTableCheckOperator,
)

SQLColumnCheckOperator(
    task_id="check_columns",
    table="my_table",
    column_mapping={
        "id": {"null_check": {"equal_to": 0}},
    },
)

SQLTableCheckOperator(
    task_id="check_table",
    table="my_table",
    checks={"row_count": {"check_statement": "COUNT(*) > 0"}},
)

Anti-Patterns

DON'T: Access Metadata DB Directly

# WRONG - Fails in Airflow 3
from airflow.settings import Session
session.query(DagModel).all()

DON'T: Use Deprecated Imports

# WRONG
from airflow.operators.dummy_operator import DummyOperator

# CORRECT
from airflow.providers.standard.operators.empty import EmptyOperator

DON'T: Use SubDAGs

# WRONG
from airflow.operators.subdag import SubDagOperator

# CORRECT - Use task groups instead
from airflow.decorators import task_group  # AF3: from airflow.sdk import task_group

DON'T: Use Deprecated Context Keys

# WRONG
execution_date = context["execution_date"]

# CORRECT
logical_date = context["dag_run"].logical_date
data_start = context["data_interval_start"]

DON'T: Hard-Code File Paths

# WRONG
open("include/data.csv")

# CORRECT - Files in dags/
import os
dag_dir = os.path.dirname(__file__)
open(os.path.join(dag_dir, "data.csv"))

# CORRECT - Files in include/
open(f"{os.getenv('AIRFLOW_HOME')}/include/data.csv")

DON'T: Use datetime.now() in Tasks

# WRONG - Not idempotent
today = datetime.today()

# CORRECT - Use execution context
@task
def process(**context):
    logical_date = context["logical_date"]
    start = context["data_interval_start"]

Assets (Airflow 3.x)

Data-driven scheduling between DAGs:

from airflow.sdk import dag, task, Asset

# Producer — declares what data this task writes
@dag(schedule="@hourly")
def extract():
    @task(outlets=[Asset("orders_raw")])
    def pull(): ...

# Consumer — triggered when asset updates
@dag(schedule=[Asset("orders_raw")])
def transform():
    @task
    def process(): ...

Outlets without inlets are valid. A task can declare outlets even if no other DAG currently uses that asset as an inlet/schedule trigger. Outlet-only assets are encouraged for lineage tracking.