Airflow Cheat Sheet

A DAG with TaskFlow

from datetime import datetime, timedelta
from airflow.sdk import dag, task          # Airflow 2.x: from airflow.decorators import dag, task

@dag(
    schedule="@daily",
    start_date=datetime(2026, 1, 1),
    catchup=False,
    default_args={"retries": 3, "retry_delay": timedelta(minutes=5)},
    tags=["orders"],
)
def orders_daily():
    @task
    def extract(ds=None):                  # ds = logical date as YYYY-MM-DD
        return f"s3://bucket/raw/dt={ds}/"

    @task
    def load(path: str):
        print("loading", path)

    load(extract())

orders_daily()

Schedules

Value	Meaning
"@daily", "@hourly"	Presets (cron shortcuts)
"0 6 * * *"	Cron: 06:00 every day
timedelta(hours=4)	Fixed frequency
None	Manual, API or externally triggered only
[Asset("s3://lake/orders/")]	Run when an upstream task updates the asset

In Airflow 3 a cron string becomes a CronTriggerTimetable: the run fires at the cron time, its logical date is that time, and its data interval is empty (data_interval_start == data_interval_end). Airflow 2 used CronDataIntervalTimetable, where the run labelled 1 March covered 1 March and started on 2 March. If your tasks rely on the data interval, pass CronTriggerTimetable(cron, timezone="UTC", interval=timedelta(days=1)) or set [scheduler] create_cron_data_intervals = True.

Select data with the run’s own dates (ds, logical_date, data_interval_start, data_interval_end), never datetime.now().

Retries

retries, retry_delay, retry_exponential_backoff (a multiplier in Airflow 3, for example 2.0; 0 disables it), max_retry_delay, execution_timeout. Raise AirflowFailException for errors a retry cannot fix. Only safe with idempotent tasks.

Sensors

Mode	Worker slot while waiting	Use for
mode="poke" (default)	Held the whole time	Waits of a few minutes
mode="reschedule"	Released between pokes	Long waits, poke_interval of a minute or more
deferrable=True	Released; the triggerer waits	Long waits at scale (needs a running airflow triggerer)

Set timeout (the default is 7 days) and soft_fail=True when a missing input should skip rather than fail. Between DAGs, prefer asset-aware scheduling over ExternalTaskSensor.

Trigger rules

all_success (default), all_done (cleanup), one_failed (alerting), none_failed, none_failed_min_one_success (join after a branch). Prefer .as_teardown() tasks for cleanup so failures still show in the run state.

CLI (Airflow 3)

airflow dags list
airflow dags test orders_daily 2026-10-01           # run a whole DAG locally for one date
airflow tasks test orders_daily extract 2026-10-01  # run one task, no state recorded
airflow backfill create --dag-id orders_daily --from-date 2026-09-01 --to-date 2026-09-30

In Airflow 2.x, backfills used airflow dags backfill -s START -e END dag_id.

What changed in Airflow 3

Item	Airflow 2.x	Airflow 3
DAG authoring imports	airflow.decorators, airflow.models	airflow.sdk
Bash, Python, Empty operators	airflow.operators.*	airflow.providers.standard.*
Schedule argument	schedule_interval (renamed in 2.4)	schedule only
Cron timetable	CronDataIntervalTimetable	CronTriggerTimetable
catchup default	True	False
Run date in templates	execution_date	logical_date (can be None for manual runs)
Data dependencies	Datasets	Assets
Late-run alerting	sla, sla_miss_callback	Deadline alerts (3.1+)
Grouping	SubDAGs or TaskGroups	TaskGroups only
Backfill	airflow dags backfill in the CLI process	airflow backfill create, run by the scheduler
Default executor	SequentialExecutor	LocalExecutor
REST API	/api/v1	/api/v2 with bearer tokens

Rules of thumb

• Orchestrate, do not process: submit heavy work to Spark or the warehouse.
• Pass file paths or table names between tasks, not data.
• Make tasks idempotent and parameterised by the data interval.
• Set catchup explicitly.