DAG.add_task doesn't pick up start_date default arg

Apache Airflow version: 2.1.0

What happened: When attempting to set the dag for a task (Operator) without the dag previously set, where the dag’s start_date is set via default_args , an error is raised: AirflowException: Task is missing the start_date parameter.

What you expected to happen: The start_date from default_args to be honored and no error to occur.

Related notes: This relates to an issue in 1.x where using the bitshift operator with Operators that don’t specify a dag (or are not within a DAG’s context), and a dag where start_date is set via default_args, the bitshift operation will error with: AirflowException: Task is missing the start_date parameter. That issue was raised several times in the past (most recently in #7996), and @ashb had a solution in #5598 , but it was put on hold since support for bitshift operators was being debated for Airflow 2.x.

How to reproduce it:

from datetime import datetime

from airflow import DAG
from airflow.operators.dummy import DummyOperator

default_args = {
    'owner': 'airflow',
    'start_date': datetime(2017, 2, 1)
}

dag = DAG('my_dag', default_args=default_args)
dummy = DummyOperator(task_id='dummy')
dag.add_task(dummy)
# # Rest of the flow
# dummy2 = DummyOperator(task_id='dummy2')
# dag.add_task(dummy2)

# dummy >> dummy2

Setting start_date directly instead of from default_args works as expected:

...
dag = DAG('my_dag', start_date=datetime(2017, 2, 1), default_args=default_args)
...

While the example is contrived, I think it’s important functionality for reusable, pre-configured Operators, such as are used in factory and composition patterns.

[uranusjr]

Personally I wish Airflow never had default_args, it’s too magical with a logical native Python approach (**kwargs). But I guess that ship sailed way back and we should make it work mroe closely to what users expect in this case.

[eladkal]

Using the DAG code as shared by the author will show broken dag message

However if using the code as:

default_args = {
    'owner': 'airflow',
    'start_date': datetime(2017, 2, 1)
}
with DAG('my_dag', default_args=default_args):
    DummyOperator(task_id='dummy')

It will work

In our examples we never show usage of dag.add_task() and I’m not sure why this is needed here? Noting also Ash comment to set the start_date on the DAG level. Since the recommended way to use context manager and this is working fine I’m closing this issue.