Tasks in DAGs with depends_on_past or task_concurrency are not being scheduled

[nathadfield]

Apache Airflow version: 1.10.13

What happened:

After performing an upgrade to v1.10.13 we noticed that tasks in some of our DAGs were not be scheduled. After a bit of investigation we discovered that by commenting out 'depends_on_past': True the issue went away.

What you expected to happen:

We think the issue might have something to do with this which was introduced to 1.10.13

[AIRFLOW-3607] Only query DB once per DAG run for TriggerRuleDep (#4751)

How to reproduce it:

1. Install Airflow v1.10.13 from pip
2. Start webserver and scheduler
3. Add the following code as a DAG
4. Switch the DAG on in the UI.

from airflow import models
from airflow.operators.dummy_operator import DummyOperator
from datetime import datetime, timedelta

default_args = {
    'owner': 'airflow',
    'start_date': datetime(2018, 10, 31),
    'depends_on_past': True,
    'retries': 3,
    'retry_delay': timedelta(minutes=5)
}

dag_name = 'my-test-dag'

with models.DAG(dag_name,
                default_args=default_args,
                schedule_interval='0 0 * * *',
                catchup=False,
                max_active_runs=5,
                ) as dag:

    test = DummyOperator(
        task_id='test'
    ) 

[kaxil]

On Master, this was fixed by #7402 & further optimised by #7503

in 1.10.13 -- this was clubbed by the following 2 commits:

• cb8d53f#diff-649fbbf224bab54417f03338c27d0fdb3c3336e53a522a13dfd9806c99f63137
• 268f1be#diff-649fbbf224bab54417f03338c27d0fdb3c3336e53a522a13dfd9806c99f63137

I will investigate this further

[kaxil]

@nathadfield has confirmed the issue does not exist on 2.0.0b3

[mthoretton]

Since v1.10.13 we also noticed, for some dags, the tasks are not being scheduled. They stay forever with a None state. Nothing in the scheduler logs (DEBUG level). Running the tasks manually work fine though. In our case, some/most of the dags have indeed depends_on_past set to true but not all of them it seems. So maybe this is something different. I will try to investigate deeper and share any relevant info.

[kaxil]

Since v1.10.13 we also noticed, for some dags, the tasks are not being scheduled. They stay forever with a None state. Nothing in the scheduler logs (DEBUG level). Running the tasks manually work fine though. In our case, some/most of the dags have indeed depends_on_past set to true but not all of them it seems. So maybe this is something different. I will try to investigate deeper and share any relevant info.

Can you check if the other DAGs (not using depends_on_past but are still stuck) have task_concurrency set? @mthoretton

[mthoretton]

I did not check absolutely all dags we have but yes, the "broken" dags either have depends_on_past or task_concurrency set. I was not aware of the 2 issues you mentionened above, I will have a look but it definitely looks related.

[kaxil]

I can confirm the bug. I was able to reproduce it with task with task_concurrency or depends_on_past with LocalExecutor and the following DAG:

from airflow import models
from airflow.operators.dummy_operator import DummyOperator
from airflow.operators.bash_operator import BashOperator
from datetime import datetime, timedelta

default_args = {
    'owner': 'airflow',
    'start_date': datetime(2018, 10, 31),
    'retries': 3,
    'retry_delay': timedelta(minutes=5)
}

dag_name = 'dag-bugcheck'

with models.DAG(dag_name,
                default_args=default_args,
                schedule_interval='0 0 * * *',
                catchup=False,
                max_active_runs=5,
                ) as dag:

    test1 = DummyOperator(
        task_id='test1',
        task_concurrency=10,
    )

    test2 = BashOperator(
        task_id='test2',
        bash_command='echo hi',
        depends_on_past=True,
    )

    test3 = BashOperator(
        task_id='test3',
        bash_command='echo hi',
    )

[kaxil]

#12663 should fix it @nathadfield @mthoretton

[nathadfield]

Nice one @kaxil! Will this force the need for a 1.10.14 then?

[kaxil]

Nice one @kaxil! Will this force the need for a 1.10.14 then?

Yup, indeed. I hope to get it out by early next week

[kaxil]

Closed by #12663