Airflow ExternalTaskSensor don't fail when External Task fails

failed_states was added in Airflow 2.0; you'd set it to ["failed"] to configure the sensor to fail the current DAG run if the monitored DAG run failed. If given a task ID, it'll monitor the task state, otherwise it monitors DAG run state.

In Airflow 1.x, unfortunately, the ExternalTaskSensor operation only compares DAG run or task state against allowed_states; as soon as the monitored DAG run or task reaches one of the allowed states, the sensor stops, and is then always marked as successful. By default, the sensor only looks for the SUCCESS state, so without a timeout it'll just keep on poking forever if the monitored DAG run has failed. If you put failed in the allowed_states list, it will still only ever mark itself as successful.

While you could use a timeout, like you I needed the sensor to fail it's own DAG run if the external DAG run failed, as if the dependencies for the next task have not been met. This requires you write your own sensor, unfortunately.

Here is my implementation; it is a simplified version of the ExternalTaskSensor() class, adapted to my simpler needs (no need to check for a specific task id or for anything other than the same execution date):

from airflow.exceptions import AirflowFailException
from airflow.models import DagRun
from airflow.sensors.base_sensor_operator import BaseSensorOperator
from airflow.utils.db import provide_session
from airflow.utils.decorators import apply_defaults
from airflow.utils.state import State

class ExternalDagrunSensor(BaseSensorOperator):
    """
    Waits for a different DAG to complete; if the dagrun has failed, this
    task fails itself as well.

    :param external_dag_id: The dag_id that contains the task you want to
        wait for
    :type external_dag_id: str
    """

    template_fields = ["external_dag_id"]
    ui_color = "#19647e"

    @apply_defaults
    def __init__(self, external_dag_id, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.external_dag_id = external_dag_id

    @provide_session
    def poke(self, context, session=None):
        dag_id, execution_date = self.external_dag_id, context["execution_date"]
        self.log.info("Poking for %s on %s ... ", dag_id, execution_date)

        state = (
            session.query(DagRun.state)
            .filter(
                DagRun.dag_id == dag_id,
                DagRun.execution_date == execution_date,
                DagRun.state.in_((State.SUCCESS, State.FAILED)),
            )
            .scalar()
        )
        if state == State.FAILED:
            raise AirflowFailException(
                f"The external DAG run {dag_id} {execution_date} has failed"
            )
        return state is not None

The base sensor implementation will call the poke() method repeatedly until it returns True (or the optional timeout was reached), and by raising AirflowFailException the task state is set to failed immediately, no retrying. It is then up to the downstream task configuration if they will be scheduled to run.

ExternalTaskSensor just pokes till some expected state is reached, it's state is not intended to be mapped with the external task state.

it defaults to [State.SUCCESS] that's why if success you don't have any problem. Adding allowed_states=[State.SUCCESS, State.failed, State.upstream_failed] To your code will at least ensure the external task has finished.

Additionally you can set a timeout to make it fail, if soft_fail = False

If you want for the sensor to FAIL if the external task failed you'll need to write your own implementation of such sensor.

For example here's how I'm checking for Last Dagrun of a Dag to match certain state

@provide_session
def poke(self, context, session=None):
    """
    Checks if latest dag_run State is in expected state else keeps polling...
    :param context:
    :param session:
    :return:
    """
    DR = DagRun
    self.log.info(
        f"Poking for {self.external_dag_id}, {self.allowed_states} -> {self.state_condition} ... "
    )
    # If state is expected to match
    if self.state_condition:
        query = session.query(DR).filter(DR.dag_id == self.external_dag_id,
                                                   DR.state.notin_(self.allowed_states))
    # If state is not expected to match
    else:
        query = session.query(DR).filter(DR.dag_id == self.external_dag_id,
                                                   DR.state.in_(self.allowed_states))
    # Filter by last_dagrun, could be max(execution_date) also but avoiding such aggregation
    # by sorting dag_run chronologically in descendent order
    query = query.order_by(DR.execution_date.desc()).first()
    session.commit()
    return not query