Airflow Version Detection

When reading a DAG with read_dag, the LLM sees the raw source code and can identify the Airflow version from imports and patterns. This guide helps with version-appropriate conversion.

Detection Strategy

Version detection uses a combination of:

1. Import patterns — Most reliable indicator
2. API usage — Decorator styles, method signatures
3. Feature presence — Datasets, Assets, TaskFlow

Version Indicators

Airflow 1.x

# 1.x-specific imports
from airflow.operators.python_operator import PythonOperator
from airflow.operators.bash_operator import BashOperator
from airflow.hooks.base_hook import BaseHook

# 1.x patterns
t1.set_downstream(t2)
t2.set_upstream(t1)

Signals:

• airflow.operators.python_operator (renamed in 2.x)
• airflow.operators.bash_operator (renamed in 2.x)
• set_downstream() / set_upstream() methods
• Missing @task / @dag decorators

Airflow 2.x

# 2.x imports
from airflow.operators.python import PythonOperator
from airflow.operators.bash import BashOperator
from airflow.decorators import dag, task
from airflow.datasets import Dataset

# 2.x patterns
@dag
def my_dag():
    @task
    def my_task():
        ...

Signals:

• airflow.decorators — TaskFlow API (2.0+)
• airflow.datasets — Dataset scheduling (2.4+)
• from airflow.operators.python import (not python_operator)
• >> / << dependency operators

Airflow 3.x

# 3.x imports
from airflow.assets import Asset
from airflow.sdk import DAG, task

# 3.x patterns
my_asset = Asset("s3://bucket/data")

@task(outlets=[my_asset])
def produce_data():
    ...

Signals:

• airflow.assets — Asset API (replaces Dataset in 3.x)
• airflow.sdk — New SDK module
• Asset class with outlets/inlets

Feature Indicators

When reading a DAG, look for these features and use lookup_concept for each:

TaskFlow API

Detected when: @dag or @task decorators present

Lookup: lookup_concept("dag-to-flow") — maps directly to Prefect @flow / @task

Datasets (2.4+)

Detected when: from airflow.datasets import Dataset

Lookup: lookup_concept("dataset-to-automation") — convert to Prefect events or triggers

Dynamic Task Mapping

Detected when: .expand() method calls

Lookup: lookup_concept("dynamic-mapping") — convert to Prefect .map()

Version-Specific Guidance

From 1.x

• Update all import paths
• Replace set_downstream/set_upstream with function call ordering
• XCom pull/push → return values and parameters
• Consider using TaskFlow patterns for cleaner code

From 2.x

• TaskFlow maps directly to Prefect @flow/@task
• Datasets → Events or triggers
• Connections → Blocks
• Pools → Concurrency limits

From 3.x

• Assets → Event-based automation
• SDK patterns → Native Prefect patterns
• Execution context → Run context

Handling Ambiguity

When version cannot be determined:

1. Multiple indicators: Most recent version wins
2. No clear indicators: Default to 2.x patterns
3. Mixed patterns: Note in conversion comments