Apache Airflow¶

Apache Airflow is an open-source platform for authoring, scheduling, and monitoring workflows as Directed Acyclic Graphs (DAGs). It is the de facto standard for data pipeline orchestration.

Architecture¶

Component	Role
Web Server	HTTP interface, user UI (default port 8080)
Scheduler	Periodically checks registered DAGs against schedule, creates DAG Runs
Worker	Executes tasks from the queue
Queue	Tasks waiting for execution (Redis, RabbitMQ)
Metastore	Stores DAG definitions, run history, task states (PostgreSQL, MySQL)

Scales horizontally by adding Workers and Web Servers.

DAG Structure (Five Blocks)¶

# 1. IMPORTS
from airflow import DAG
from airflow.operators.python import PythonOperator
from datetime import datetime, timedelta

# 2. TASK FUNCTIONS
def extract(): pass
def transform(): pass
def load(): pass

# 3. DEFAULT ARGS + DAG
default_args = {
    'owner': 'data_team',
    'depends_on_past': False,
    'retries': 2,
    'retry_delay': timedelta(minutes=5),
    'start_date': datetime(2024, 1, 1),
    'email_on_failure': True,
    'retry_exponential_backoff': True,
    'max_retry_delay': timedelta(minutes=30),
}

dag = DAG('etl_pipeline', default_args=default_args,
          schedule_interval='0 12 * * *', catchup=False)

# 4. TASK INITIALIZATION
t1 = PythonOperator(task_id='extract', python_callable=extract, dag=dag)
t2 = PythonOperator(task_id='transform', python_callable=transform, dag=dag)
t3 = PythonOperator(task_id='load', python_callable=load, dag=dag)

# 5. DEPENDENCIES
t1 >> t2 >> t3

TaskFlow API (Airflow 2.0+)¶

from airflow.decorators import dag, task

@dag(schedule_interval='@daily', start_date=days_ago(1), catchup=False)
def etl_pipeline():
    @task(retries=3)
    def extract():
        return requests.get(URL).content

    @task()
    def transform(raw_data):
        return pd.read_csv(StringIO(raw_data)).to_csv(index=False)

    @task()
    def load(data):
        print(f"Loading {len(data)} bytes")

    raw = extract()
    table = transform(raw)
    load(table)

etl_pipeline()

Advantages: XCom handled automatically, dependencies inferred from call chain, per-task retry overrides.

Schedule Options¶

Method	Example	Use Case
Cron	`"0 2 * * *"`	Fixed calendar-based
Presets	`@daily`, `@hourly`, `@weekly`	Common intervals
timedelta	`timedelta(hours=2)`	Fixed interval from last run
`@once`	Single execution	Manual triggers
`None`	No schedule	Manual only

Operators and Sensors¶

Type	Examples
PythonOperator	Run Python functions
BashOperator	Run shell commands
DummyOperator	DAG flow structuring, join points
S3KeySensor	Wait for file in S3
ExternalTaskSensor	Wait for task in another DAG
HttpSensor	Wait for HTTP endpoint success

Trigger Rules¶

Rule	Behavior
`ALL_SUCCESS`	Default - all upstream succeeded
`ONE_SUCCESS`	At least one upstream succeeded
`ALL_FAILED`	All upstream failed
`ALL_DONE`	All upstream completed (any status)

XCom (Cross-Communication)¶

def multiply(**context):
    value = context['task_instance'].xcom_pull(
        task_ids='load_task', key='return_value')
    return value * value  # auto-pushed to XCom

# TaskFlow API: XCom is implicit via function params/returns

Key Parameters¶

catchup=True/False - whether to backfill missed runs
max_active_runs - limit concurrent DAG runs
execution_timeout - kill task if too long
depends_on_past - task waits for previous run's success
on_failure_callback - alert function on failure

Built-in Context Variables¶

ds - execution date (YYYY-MM-DD string)
execution_date - execution datetime object
dag - DAG object
task_instance / ti - current TaskInstance

Gotchas¶

schedule_interval defines interval between runs, not run time. @daily with start_date=Jan 1 first runs on Jan 2
catchup=True (default) will backfill all missed intervals - can flood the system
provide_context=True required for XCom in classic API; TaskFlow handles automatically
When upstream fails with ALL_SUCCESS, downstream shows "upstream_failed" (not "failed")
DAG files go in configured dags/ directory; deploy by pushing to Git