tasks with `none_failed` trigger rule do not run when all upstream tasks are skipped

Apache Airflow version: 1.10.12 Kubernetes version (if you are using kubernetes) (use kubectl version): v1.16.15-gke.6000

Environment:

• Cloud provider or hardware configuration: Google Kubernetes Engine
• OS (e.g. from /etc/os-release): “Container-Optimized OS”

BUILD_ID=12371.1088.0
NAME="Container-Optimized OS"
KERNEL_COMMIT_ID=52bdab9330bdd9e50dc967f8aa850829921ca217
GOOGLE_CRASH_ID=Lakitu
VERSION_ID=77
BUG_REPORT_URL="https://cloud.google.com/container-optimized-os/docs/resources/support-policy#contact_us"
PRETTY_NAME="Container-Optimized OS from Google"
VERSION=77
GOOGLE_METRICS_PRODUCT_ID=26
HOME_URL="https://cloud.google.com/container-optimized-os/docs"
ID=cos

• Kernel (e.g. uname -a): Linux gke-services-1-default-pool-3ef08c09-v95l 4.19.112+ #1 SMP Sat Oct 10 13:45:37 PDT 2020 x86_64 Intel(R) Xeon(R) CPU @ 2.30GHz GenuineIntel GNU/Linux
• Install tools: helm, chart airflow-7.16.0
• Others:

What happened:

Even after https://github.com/apache/airflow/pull/7464, we are finding that tasks with the none_failed trigger are still being skipped when their direct upstream is skipped.

In a simple three-task DAG where, the first step is the GoogleCloudStoragePrefixSensor, followed by a processing task and ending with a heartbeat check operator:

check_for_late_data >> run_statekeeper >> passive_check

The passive_check task is configured with NONE_FAILED:

passive_check = PassiveCheckOperator(
    task_id="passive_check",
    dag=dag,
    trigger_rule=TriggerRule.NONE_FAILED)

The GCS sensor operator exits thusly if it finds no keys:

[2021-02-12 00:32:18,130] {taskinstance.py:1025} INFO - Marking task as SKIPPED.dag_id=pipeline_v1, task_id=check_for_late_data, execution_date=20210211T003000, start_date=20210212T003017, end_date=
[2021-02-12 00:32:18,130] {taskinstance.py:1070} INFO - Marking task as SUCCESS.dag_id=pipeline_v1, task_id=check_for_late_data, execution_date=20210211T003000, start_date=20210212T003017, end_date=20210212T003218

The intermediate step is also skipped as is intended (it uses the default trigger rule). But the final step is skipped as well, which should not happen:

The same thing happens if we put a dummy shared start task upstream:

What you expected to happen:

The “passive_check” task should have run, because its trigger is none_failed and no tasks upstream of it have failed, they have only been skipped.

As this was allegedly fixed in https://github.com/apache/airflow/pull/7464, I suspect that either something has regressed since 1.10.10 or there is a corner case not yet caught.

How to reproduce it:

The following DAG reproduces the issue (presuming that you have a working Google Cloud Platform connection and a GCS bucket which the DAG can read):

import datetime
from airflow.models import DAG
from airflow.operators.dummy_operator import DummyOperator
from airflow.contrib.sensors.gcs_sensor import GoogleCloudStoragePrefixSensor

SOURCE_FILE_BUCKET =  "GCS_BUCKET_NAME"  # replace this with a GCS bucket under your control
SOURCE_FILE_PREFIX = "nofileshere/"  # make sure there are zero keys under this prefix

with DAG(
    dag_id="simple_skip",
    start_date=datetime.datetime(2021, 2, 19),
    schedule_interval=None,
) as dag:
    find_no_data = GoogleCloudStoragePrefixSensor(
        dag=dag,
        task_id="find_no_data",
        soft_fail=True,
        timeout=60 * 2,
        bucket=SOURCE_FILE_BUCKET,
        prefix=SOURCE_FILE_PREFIX,
    )
    step_1 = DummyOperator(task_id="step_1", dag=dag)
    step_2 = DummyOperator(task_id="step_2", dag=dag, trigger_rule="none_failed")
    find_no_data >> step_1 >> step_2

[josh-fell]

I can confirm this is fixed in Airflow 2.1.0 and Airflow 2.2.0.

DAG code

from datetime import datetime

from airflow import DAG
from airflow.decorators import task
from airflow.exceptions import AirflowSkipException
from airflow.operators.dummy import DummyOperator
from airflow.utils.trigger_rule import TriggerRule


@task
def task1():
    raise AirflowSkipException

@task
def task2():
    raise AirflowSkipException

@task
def task3():
    raise AirflowSkipException


with DAG(
    dag_id="__test__",
    start_date=datetime(2021, 1, 1),
    schedule_interval="0 * * * *",
    tags=["test"],
    catchup=False,
) as dag:
    end = DummyOperator(task_id="end", trigger_rule=TriggerRule.NONE_FAILED)

    task1() >> task2() >> end
    task3() >> end

Sample DAG run

[n-oden]

As it turns out, this has been fixed between 1.10.12 and HEAD, but has not been released yet.

The issue was in https://github.com/apache/airflow/blob/1.10.12/airflow/sensors/base_sensor_operator.py#L112-L114:

                if self.soft_fail and not context['ti'].is_eligible_to_retry():
                    self._do_skip_downstream_tasks(context)
                    raise AirflowSkipException('Snap. Time is **OUT.')**

Any sensor that inherited BaseSensorOperator and had soft_fail set to True would raise AirflowSkipExcepton but only after manually skipping all downstream tasks in _do_skip_downstream_tasks():

    def _do_skip_downstream_tasks(self, context):
        downstream_tasks = context['task'].get_flat_relatives(upstream=False)
        self.log.debug("Downstream task_ids %s", downstream_tasks)
        if downstream_tasks:
            self.skip(context['dag_run'], context['ti'].execution_date, downstream_tasks)

This appears to have been fixed in HEAD, where in https://github.com/apache/airflow/blob/master/airflow/sensors/base.py#L233-L236 just the exception is raised and the _do_skip_downstream_tasks() function no longer exists.

And as a side note, if you look at https://github.com/apache/airflow/blob/master/airflow/models/taskinstance.py#L1142-L1157, the _run_raw_task() method is primarily a very large try/except block that attempts to run the task and then handle each possible exception in turn. But the exception handler for AirflowSkipException does not return, which means that the flow falls out of the try/except block and continues on L1182 with the task being marked success. You can actually see this in our log snippet above:

[2021-02-12 00:32:18,130] {taskinstance.py:1025} INFO - Marking task as SKIPPED.dag_id=pipeline_v1, task_id=check_for_late_data, execution_date=20210211T003000, start_date=20210212T003017, end_date=
[2021-02-12 00:32:18,130] {taskinstance.py:1070} INFO - Marking task as SUCCESS.dag_id=pipeline_v1, task_id=check_for_late_data, execution_date=20210211T003000, start_date=20210212T003017, end_date=20210212T003218

I suspect that there should be a raise or a return at the end of that block?