Removed mapped tasks block all_success trigger rule

[tanelk]

Apache Airflow version

2.6.3

What happened

When rerunning a DAG run with dynamically mapped tasks and the number of mapped task instances degreases, then downstream tasks with all_success trigger rule (likely some others as well) will not get scheduled, because removed status is not considered to be successful. When a regular task gets removed, then this does not happen, because it will get removed from the DAG structure.

Task dependency message
Task's trigger rule 'all_success' requires all upstream tasks to have succeeded, but found 2 non-success(es). upstream_states=_UpstreamTIStates(success=22, skipped=0, failed=0, upstream_failed=0, removed=2, done=24), upstream_task_ids={'upstream_task'}

Previously you could forcefully run it with the "run" button in UI, but it has been removed in one of the resent releases.

Luckily I could delete the removed TIs and then the downstream tasks got rescheduled, but this is not "scalable".

What you think should happen instead

Removed TIs should not block any trigger rules from getting executed.

How to reproduce

Operating System

Versions of Apache Airflow Providers

No response

Deployment

Virtualenv installation

Deployment details

No response

Anything else

No response

Are you willing to submit PR?

• Yes I am willing to submit a PR!

Code of Conduct

• I agree to follow this project's Code of Conduct

[dzhigimont]

@tanelk Could you please provide an example of DAG?

[RNHTTR]

Can you please provide reproduction steps?

Also, what do you mean by the following:

the number of mapped task instances decreases

How are the number of mapped task instances decreasing during a DAG run?

[tanelk]

Example DAG

from airflow import DAG
from airflow.decorators import task
from airflow.operators.empty import EmptyOperator

with DAG(
    'removed_mapped_tasks',
    schedule='@daily',
    start_date=pendulum.DateTime(2023, 8, 7),
) as dag:

    @task
    def gen_elements():
        return [1, 2, 3]

    @task
    def mapped_task(element):
        return element * 2

    mapped_task.expand(element=gen_elements()) >> EmptyOperator(task_id='end')

Let it complete and then return one less element from the gen_elements task. Then clear the last DAG run.

The end task will not get scheduled because Task's trigger rule 'all_success' requires all upstream tasks to have succeeded, but found 1 non-success(es). upstream_states=_UpstreamTIStates(success=2, skipped=0, failed=0, upstream_failed=0, removed=1, done=3), upstream_task_ids={'mapped_task'}

In this very simple DAG, the run will be failed with scheduler | [2023-08-09T13:49:56.448+0300] {dagrun.py:651} ERROR - Task deadlock (no runnable tasks); marking run <DagRun removed_mapped_tasks @ 2023-08-06 21:00:00+00:00: scheduled__2023-08-06T21:00:00+00:00, state:running, queued_at: 2023-08-09 10:49:44.110567+00:00. externally triggered: False> failed

On more complex structures the deadlock detection might not kick in.