bperson opened a new pull request #14476: URL: https://github.com/apache/airflow/pull/14476
<!-- Thank you for contributing! Please make sure that your code changes are covered with tests. And in case of new features or big changes remember to adjust the documentation. Feel free to ping committers for the review! In case of existing issue, reference it using one of the following: closes: #ISSUE related: #ISSUE How to write a good git commit message: http://chris.beams.io/posts/git-commit/ --> Currently the scheduler can be stuck in a loop if the query tweaked in this PR continually returns a set of task instances that aren't schedule-able because all their respective pools are full. This is even more apparent when your `max_tis` gets low with the various `min`s in the scheduler main loop code path ( either because your overall pool open slots are low, your parallelism is low, or you've set `max_tis_per_query` too low ). The unit test reproduces the issue ( though I'm still not convinced it's the best way to show it, I've included further down a "setup", I guess this should be thrown into an integration test instead? ): The test setup idea is to fill the set of schedule-able TIs returned by that query full of TIs that will starve a first pool from open slots, and then add another set of schedule-able TIs that should be easy to schedule and execute because their execution pool is much bigger. Without filtering out TIs from the first pool ( that will be further discarded in that code path ) we end up with a scheduler stuck until the first pool gets through its work, this starves the second pool for no reason other than the scheduler never getting to them: Using a minimal "real-life" example ( you need to create a `starving_pool` with only a couple of slots ) ``` [2021-02-26 08:56:43,604] {scheduler_job.py:950} INFO - 2 tasks up for execution: <TaskInstance: starving_pool.sleeper 2021-02-26 03:20:00+00:00 [scheduled]> <TaskInstance: starving_pool.sleeper 2021-02-26 03:10:00+00:00 [scheduled]> [2021-02-26 08:56:43,606] {scheduler_job.py:979} INFO - Figuring out tasks to run in Pool(name=starving_pool) with 0 open slots and 2 task instances ready to be queued [2021-02-26 08:56:43,606] {scheduler_job.py:994} INFO - Not scheduling since there are 0 open slots in pool starving_pool [2021-02-26 08:56:43,606] {scheduler_job.py:1072} INFO - Setting the following tasks to queued state: [2021-02-26 08:56:44,757] {scheduler_job.py:950} INFO - 2 tasks up for execution: <TaskInstance: starving_pool.sleeper 2021-02-26 03:20:00+00:00 [scheduled]> <TaskInstance: starving_pool.sleeper 2021-02-26 03:10:00+00:00 [scheduled]> [2021-02-26 08:56:44,759] {scheduler_job.py:979} INFO - Figuring out tasks to run in Pool(name=starving_pool) with 0 open slots and 2 task instances ready to be queued [2021-02-26 08:56:44,759] {scheduler_job.py:994} INFO - Not scheduling since there are 0 open slots in pool starving_pool [2021-02-26 08:56:44,759] {scheduler_job.py:1072} INFO - Setting the following tasks to queued state: [2021-02-26 08:56:45,068] {scheduler_job.py:950} INFO - 2 tasks up for execution: <TaskInstance: starving_pool.sleeper 2021-02-26 03:20:00+00:00 [scheduled]> <TaskInstance: starving_pool.sleeper 2021-02-26 03:10:00+00:00 [scheduled]> [2021-02-26 08:56:45,070] {scheduler_job.py:979} INFO - Figuring out tasks to run in Pool(name=starving_pool) with 0 open slots and 2 task instances ready to be queued [2021-02-26 08:56:45,070] {scheduler_job.py:994} INFO - Not scheduling since there are 0 open slots in pool starving_pool [2021-02-26 08:56:45,070] {scheduler_job.py:1072} INFO - Setting the following tasks to queued state: [2021-02-26 08:56:45,309] {scheduler_job.py:950} INFO - 2 tasks up for execution: <TaskInstance: starving_pool.sleeper 2021-02-26 03:20:00+00:00 [scheduled]> <TaskInstance: starving_pool.sleeper 2021-02-26 03:10:00+00:00 [scheduled]> [2021-02-26 08:56:45,311] {scheduler_job.py:979} INFO - Figuring out tasks to run in Pool(name=starving_pool) with 0 open slots and 2 task instances ready to be queued [2021-02-26 08:56:45,311] {scheduler_job.py:994} INFO - Not scheduling since there are 0 open slots in pool starving_pool ``` `starved_dag.py` ``` """DAG starved even though its TIs are executable in a non starved pool ( default pool )""" from datetime import datetime, timedelta from airflow.models.dag import DAG from airflow.operators.bash import BashOperator from airflow.utils.dates import days_ago DAG_NAME = 'starved_dag' default_args = {'owner': 'airflow', 'start_date': days_ago(0), 'dagrun_timeout': timedelta(minutes=6)} dag = DAG(DAG_NAME, schedule_interval='*/10 * * * *', default_args=default_args) echoer = BashOperator( task_id='echoer', bash_command='echo "bonjour"', dag=dag, ) if __name__ == "__main__": dag.cli() ``` `starving_pool.py` ``` """DAG to starve schedulable TIs query""" from datetime import datetime, timedelta from airflow.models.dag import DAG from airflow.operators.bash import BashOperator from airflow.utils.dates import days_ago DAG_NAME = 'starving_pool' default_args = {'owner': 'airflow', 'start_date': days_ago(0), 'dagrun_timeout': timedelta(minutes=6)} dag = DAG(DAG_NAME, schedule_interval='*/10 * * * *', default_args=default_args) sleeper = BashOperator( task_id='sleeper', bash_command='sleep 300', dag=dag, pool='starving_pool' ) if __name__ == "__main__": dag.cli() ``` We hit this issue in our production stack with real life DAGs. To bypass it, we're currently running with an insanely big unused pool ( `999999` slots :D ), an insanely big `max_tis_per_query`, and an insanely big `parallelism` ( even though we don't have the actual workers behind it ) to make sure that the number of TIs returned by the tweaked query will always be bigger than the maximum number of TIs stuck in starved pools at any point in time. --- **^ Add meaningful description above** Read the **[Pull Request Guidelines](https://github.com/apache/airflow/blob/master/CONTRIBUTING.rst#pull-request-guidelines)** for more information. In case of fundamental code change, Airflow Improvement Proposal ([AIP](https://cwiki.apache.org/confluence/display/AIRFLOW/Airflow+Improvements+Proposals)) is needed. In case of a new dependency, check compliance with the [ASF 3rd Party License Policy](https://www.apache.org/legal/resolved.html#category-x). In case of backwards incompatible changes please leave a note in [UPDATING.md](https://github.com/apache/airflow/blob/master/UPDATING.md). ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected]
