[
https://issues.apache.org/jira/browse/AIRFLOW-6388?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
t oo updated AIRFLOW-6388:
--------------------------
Description:
Spark jobs can often take many minutes (or even hours) to complete.
The spark submit operator submits a job to a spark cluster, then continually
polls its status until it detects the spark job has ended. This means it could
be consuming a 'slot' (ie parallelism, dag_concurrency,
max_active_dag_runs_per_dag, non_pooled_task_slot_count) for hours when it is
not 'doing' anything but polling for status.
https://github.com/apache/airflow/pull/6909#discussion_r361838225 suggested it
should move to a poke/reschedule model.
Another thing to note is that in cluster mode a spark-submit made to a 'full'
spark cluster will sit in WAITING state on spark side until some cores/memory
is freed, then the driver/app can go into RUNNING
"This actually means occupy worker and do nothing for n seconds is it not?
It was OK when it was 1 second but users may set it to even 5 min without
realising that it occupys the worker.
My comment here is more of a concern rather than an action to do.
Should this work by occupying the worker "indefinitely" or can it be something
like the sensors with (poke/reschedule)?"
was:
Spark jobs can often take many minutes (or even hours) to complete. The spark
submit operator submits a job to a spark cluster, then continually polls its
status until it detects the spark job has ended. This means it could be
consuming a 'slot' (ie parallelism, dag_concurrency,
max_active_dag_runs_per_dag, non_pooled_task_slot_count) for hours when it is
not 'doing' anything but polling for status.
https://github.com/apache/airflow/pull/6909#discussion_r361838225 suggested it
should move to a poke/reschedule model.
"This actually means occupy worker and do nothing for n seconds is it not?
It was OK when it was 1 second but users may set it to even 5 min without
realising that it occupys the worker.
My comment here is more of a concern rather than an action to do.
Should this work by occupying the worker "indefinitely" or can it be something
like the sensors with (poke/reschedule)?"
> SparkSubmitOperator polling should not 'consume' a slot
> -------------------------------------------------------
>
> Key: AIRFLOW-6388
> URL: https://issues.apache.org/jira/browse/AIRFLOW-6388
> Project: Apache Airflow
> Issue Type: Improvement
> Components: dependencies, scheduler
> Affects Versions: 1.10.3
> Reporter: t oo
> Priority: Minor
>
> Spark jobs can often take many minutes (or even hours) to complete.
> The spark submit operator submits a job to a spark cluster, then continually
> polls its status until it detects the spark job has ended. This means it
> could be consuming a 'slot' (ie parallelism, dag_concurrency,
> max_active_dag_runs_per_dag, non_pooled_task_slot_count) for hours when it is
> not 'doing' anything but polling for status.
> https://github.com/apache/airflow/pull/6909#discussion_r361838225 suggested
> it should move to a poke/reschedule model.
> Another thing to note is that in cluster mode a spark-submit made to a 'full'
> spark cluster will sit in WAITING state on spark side until some cores/memory
> is freed, then the driver/app can go into RUNNING
> "This actually means occupy worker and do nothing for n seconds is it not?
> It was OK when it was 1 second but users may set it to even 5 min without
> realising that it occupys the worker.
> My comment here is more of a concern rather than an action to do.
> Should this work by occupying the worker "indefinitely" or can it be
> something like the sensors with (poke/reschedule)?"
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
