---

name: airflow description: "Airflow DAG patterns, KubernetesPodOperator, and debugging. Use on 'dag', 'airflow', 'task', 'operator', 'KPO', 'scheduler', 'XCom'."

Airflow Skill

Minimal, production-grade Airflow patterns for Airflow 2 and 3.

Version Detection (Must Run First)

Before proposing any changes, confirm the Airflow version:

Detection order (use first available):

1. Project dependency pins (requirements.txt, constraints.txt, pyproject.toml)
2. Deployed image/app version (Helm values, image tags)
3. Runtime confirmation: airflow version inside scheduler/webserver pod

Hard rules:

• Do not mix constructs between Airflow major/minor versions
• Always validate guidance against the project's current Airflow version pin
• If you cannot determine the version, stop and ask one focused question

Docs rule:

• Prefer https://airflow.apache.org/docs/apache-airflow/<major>.<minor>.*
• Avoid /stable/ docs unless explicitly on current stable release

Version Support

This skill supports both Airflow 2.x and Airflow 3.x. Key differences:

Airflow 3.x Changes

• New import namespace: Use from airflow.sdk import DAG, task instead of airflow.models
• Assets replace Datasets: Dataset → Asset, DatasetEvent → AssetEvent
• No metadata DB access in tasks: Use Airflow REST API or context instead
• schedule_interval removed: Use unified schedule parameter
• catchup=False by default: Explicit opt-in for backfills
• logical_date=None for manual/asset triggers: No data interval for ad-hoc runs
• Standard operators moved: PythonOperator, BashOperator now in apache-airflow-providers-standard
• Removed: SubDAGs, SLAs, pickling, execution_date context variable

When Writing DAGs

• Airflow 2: Use legacy imports (airflow.models, airflow.decorators)
• Airflow 3: Use airflow.sdk imports for forward compatibility
• Check version with from airflow import __version__

When to Use

• Writing new DAGs
• Debugging task failures
• Optimizing scheduler performance
• Configuring KubernetesPodOperator
• Managing connections and variables

Python's Zen Applied to DAGs

CopySimple is better than complex.        -> Use TaskFlow over classic operators
Explicit is better than implicit.     -> Name tasks clearly, document dependencies
Flat is better than nested.           -> Avoid deep task groups unless necessary
Sparse is better than dense.          -> One DAG per file, focused responsibility
Errors should never pass silently.    -> Always set on_failure_callback

DAG Skeleton

Airflow 3.x (Recommended)

Copy"""One-line description of what this DAG does."""
from datetime import datetime
from airflow.sdk import DAG, task

with DAG(
    dag_id="my_dag",
    start_date=datetime(2024, 1, 1),
    schedule="@daily",
    catchup=False,  # Default in 3.x but explicit is better
    tags=["team-name"],
    default_args={"owner": "team", "retries": 1},
) as dag:

    @task
    def my_task() -> dict:
        return {"status": "done"}

    my_task()

Airflow 2.x (Legacy)

Copy"""One-line description of what this DAG does."""
from datetime import datetime
from airflow import DAG
from airflow.decorators import task

with DAG(
    dag_id="my_dag",
    start_date=datetime(2024, 1, 1),
    schedule_interval="@daily",  # 2.x uses schedule_interval
    catchup=False,  # Must set explicitly in 2.x
    tags=["team-name"],
    default_args={"owner": "team", "retries": 1},
) as dag:

    @task
    def my_task() -> dict:
        return {"status": "done"}

    my_task()

KubernetesPodOperator

Airflow 3.x

Copyfrom airflow.providers.cncf.kubernetes.operators.pod import KubernetesPodOperator

KubernetesPodOperator(
    task_id="job",
    image="myimage:v1.0.0",
    cmds=["python", "run.py"],
    namespace="airflow",
    get_logs=True,
    is_delete_operator_pod=True,
)

Airflow 2.x

Same as Airflow 3.x - no breaking changes to KubernetesPodOperator parameters.

With Resources & IRSA (Both Versions)

Copyfrom kubernetes.client import V1ResourceRequirements

KubernetesPodOperator(
    task_id="job",
    image="myimage:v1.0.0",
    namespace="airflow",
    service_account_name="my-irsa-sa",
    container_resources=V1ResourceRequirements(
        requests={"memory": "256Mi", "cpu": "100m"},
        limits={"memory": "512Mi", "cpu": "200m"},
    ),
    get_logs=True,
    is_delete_operator_pod=True,
)

Common Operators

Note: Airflow 3.x moved standard operators (PythonOperator, BashOperator, EmailOperator, SimpleHttpOperator) to apache-airflow-providers-standard package.

XCom Patterns

Copy# Return references, not data
@task
def extract() -> str:
    s3.upload(data, "s3://bucket/output.parquet")
    return "s3://bucket/output.parquet"

@task
def transform(path: str) -> str:
    data = s3.download(path)
    # ...
    return "s3://bucket/transformed.parquet"

transform(extract())

Version Notes:

• Airflow 3.x: xcom_pull(key="key") requires task_ids parameter (no more implicit pulls)
• Airflow 2.x: xcom_pull() without task_ids allowed but ambiguous (avoid)

Asset-Based Scheduling (Airflow 3.x)

Airflow 3.x renames Datasets to Assets and enhances event-driven scheduling.

Copyfrom airflow.sdk import DAG, task, Asset

# Define assets
raw_data = Asset("s3://bucket/raw/data.parquet")
clean_data = Asset("s3://bucket/clean/data.parquet")

# Producer DAG
with DAG(dag_id="producer", schedule="@daily") as producer_dag:
    @task(outlets=[raw_data])
    def extract():
        # Produces raw_data asset
        return {"status": "done"}
    
    extract()

# Consumer DAG (triggered by asset)
with DAG(dag_id="consumer", schedule=[raw_data]) as consumer_dag:
    @task(inlets=[raw_data], outlets=[clean_data])
    def transform():
        # Consumes raw_data, produces clean_data
        return {"status": "done"}
    
    transform()

Airflow 2.x equivalent: Use Dataset instead of Asset (same pattern).

Key differences:

• 3.x: from airflow.sdk import Asset
• 2.x: from airflow.datasets import Dataset
• Context variable: triggering_asset_events (3.x) vs triggering_dataset_events (2.x)

Anti-Patterns (What to Hunt)

Critical (Both Versions)

Copy# Top-level code (runs on every scheduler heartbeat)
import pandas as pd
df = pd.read_csv("data.csv")  # RUNS AT PARSE TIME

# Move into task
@task
def process():
    import pandas as pd
    df = pd.read_csv("data.csv")

Copy# Large XCom payloads
@task
def get_data():
    return huge_dataframe.to_dict()  # Stored in metadata DB!

# Use external storage
@task
def get_data():
    s3.upload(data, "s3://bucket/data.parquet")
    return "s3://bucket/data.parquet"  # Return reference only

Copy# Hardcoded connections
conn = psycopg2.connect(host="prod-db.example.com", password="secret")

# Use Airflow Connections
from airflow.hooks.postgres_hook import PostgresHook
hook = PostgresHook(postgres_conn_id="my_postgres")

Copy# Dynamic unbounded tasks
for i in range(get_count_from_db()):  # Unknown at parse time!
    task(i)

# Use expand() for dynamic mapping with bounds
@task
def get_items():
    return [1, 2, 3]  # Bounded list

@task
def process(item):
    pass

process.expand(item=get_items())

Version-Specific Anti-Patterns

Airflow 3.x:

Copy# DON'T: Access metadata DB in tasks
from airflow.models import DagRun
dag_runs = DagRun.query.all()  # FAILS - no DB access

# DO: Use Airflow REST API or context
from airflow import __version__
# Use requests to call Airflow API

Copy# DON'T: Use execution_date (removed)
def my_task(**context):
    date = context["execution_date"]  # KeyError in 3.x

# DO: Use logical_date (or handle None for manual triggers)
def my_task(**context):
    date = context["dag_run"].logical_date  # May be None

Copy# DON'T: Use schedule_interval (removed)
DAG(dag_id="my_dag", schedule_interval="@daily")  # Fails in 3.x

# DO: Use schedule
DAG(dag_id="my_dag", schedule="@daily")

Airflow 2.x:

Copy# DON'T: Use deprecated imports (still work but warn)
from airflow.operators.python import PythonOperator  # Deprecated

# DO: Start using provider imports for 3.x readiness
from airflow.providers.standard.operators.python import PythonOperator

Performance (Both Versions)

Copy# Heavy imports at top
import tensorflow as tf  # Slow import, every heartbeat

# Import inside task
@task
def train():
    import tensorflow as tf

Debugging Flow

Version note: CLI commands same in 2.x and 3.x, but 3.x has airflow api-server instead of airflow webserver.

1. DAG Not Appearing

Copy# Check for import errors
airflow dags list-import-errors

# Validate DAG parsing/import via Airflow
airflow dags list

# Optional syntax/import sanity check (not execution)
python dags/my_dag.py

# Check scheduler logs
kubectl logs -l component=scheduler -n airflow --tail=100

2. Task Failing

Copy# Get task logs
airflow tasks logs <dag_id> <task_id> <execution_date>

# Test task locally
airflow tasks test <dag_id> <task_id> <execution_date>

# For KPO: check pod logs
kubectl logs <pod-name> -n airflow

3. Task Stuck

Copy# Check task state
airflow tasks state <dag_id> <task_id> <execution_date>

# Check for zombie tasks
airflow tasks clear <dag_id> -t <task_id> -s <start> -e <end>

# Check executor capacity
kubectl get pods -n airflow -l component=worker

4. Scheduler Slow

Copy# Check parse times
airflow dags report

# Find slow DAGs (> 1s parse time is bad)
# Optimize: remove top-level imports, reduce file count

Quick Commands

Copy# Validate DAG
airflow dags test <dag_id> <execution_date>

# Trigger DAG
airflow dags trigger <dag_id>

# Backfill
airflow dags backfill <dag_id> -s <start> -e <end>

# Clear tasks for re-run
airflow tasks clear <dag_id> -s <start> -e <end>

# List DAGs
airflow dags list

# Show DAG structure
airflow dags show <dag_id>

Response Format

When creating/modifying DAGs:

CopyDAG: <dag_id>
Schedule: <schedule>
Tasks: <task1> -> <task2> -> <task3>
Dependencies: <new providers needed>

When debugging:

CopySymptom: <what's happening>
Root cause: <why>
Fix: <action>

Minimalism Checklist

Before adding DAG code:

• Can this be a sensor waiting for data instead of polling?
• Can this use existing operators instead of PythonOperator?
• Can this use @task decorator instead of classic operator?
• Is XCom payload a reference (path/URI) not data?
• Are all imports inside tasks (not top-level)?
• Is catchup=False if backfill not needed?

Connections & Variables

Airflow 3.x:

Copy# Get connection (Task SDK)
from airflow.sdk import Connection
conn = Connection.get("my_conn")

# Get variable (Task SDK)
from airflow.sdk import Variable
val = Variable.get("my_var")

# Get secret (if Secrets Backend configured)
val = Variable.get("my_secret")  # Fetches from Secrets Manager

Airflow 2.x:

Copy# Get connection
from airflow.hooks.base import BaseHook
conn = BaseHook.get_connection("my_conn")

# Get variable
from airflow.models import Variable
val = Variable.get("my_var")

# Get secret (if Secrets Backend configured)
val = Variable.get("my_secret")  # Fetches from Secrets Manager

Task Dependencies

Copy# Chain
task1 >> task2 >> task3

# Fan out
task1 >> [task2, task3]

# Fan in
[task1, task2] >> task3

# TaskFlow (implicit)
result = task2(task1())

Version notes: Dependency syntax identical across versions.

Migration Guide (2.x → 3.x)

High Priority

1. Update imports: airflow.sdk instead of airflow.models, airflow.decorators
2. Replace schedule_interval with schedule
3. Replace execution_date with dag_run.logical_date (handle None for manual triggers)
4. Update Dataset → Asset references
5. Remove DB access from task code (use Airflow API instead)
6. Fix xcom_pull(): Always specify task_ids parameter

Medium Priority

1. Update operator imports: Move to apache-airflow-providers-standard package
2. Set catchup=False explicitly (if you rely on current behavior)
3. Remove SubDAGs: Replace with TaskGroups
4. Remove SLA callbacks: Implement custom alerting

Low Priority

1. Review deprecated config options with airflow config lint
2. Use ruff check --select AIR30 --preview to find migration issues
3. Test in Airflow 2.10+ before upgrading to 3.x

Quick Version Check

Copyfrom airflow import __version__

if __version__.startswith("3"):
    from airflow.sdk import DAG, task
else:
    from airflow import DAG
    from airflow.decorators import task