[
https://issues.apache.org/jira/browse/SPARK-19606?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16076510#comment-16076510
]
Pascal GILLET edited comment on SPARK-19606 at 9/26/17 10:27 AM:
-----------------------------------------------------------------
+1 but through 'spark.mesos.dispatcher.driverDefault.spark.mesos.constraints'!
I tested the patch and it works well!
As stated originally, the 'spark.mesos.constraints' property is ignored by the
Spark Dispatcher.
As a consequence, the Mesos slave where the Spark driver is running does not
comply to the given Mesos constraints, but on the other hand, the Mesos
constraints are well applied for the Spark executors (without the need to patch
anything).
*BUT* we do not want necessarily apply the same Mesos constraints for the
driver and executors.
For instance, we may need to run Spark drivers and executors on 2 exclusive
types of Mesos slaves:
- The dispatcher is given Mesos resources only for drivers
- Once a driver is launched, it becomes a Mesos framework itself and is
responsible for reserving resources for its executors
- If we schedule too many jobs on a Mesos cluster through the dispatcher, the
whole cluster is allocated for the drivers and there are no more resources
available for the executors. A driver may be launched but it may be waiting for
resources for its executors infinitely, which leads to a congestion then to a
dead-lock situation.
- A solution to work around this problem is to *not* mix the drivers and
executors on the same machines by passing different Mesos constraints for the
driver and for executors.
The 'spark.mesos.constraints' property still apply for executors. As for the
drivers, the 'spark.mesos.dispatcher.driverDefault.[PropertyName]' generic
property seems ideal. By definition, it allows to "_set default properties for
drivers submitted through the dispatcher_".
I propose to revise this patch and to use
'spark.mesos.dispatcher.driverDefault.spark.mesos.constraints' instead of
'spark.mesos.constraints'.
What do you guys think?
was (Author: pgillet):
+1 but through 'spark.mesos.dispatcher.driverDefault.spark.mesos.constraints'!
I tested the patch and it works well!
As stated originally, the 'spark.mesos.constraints' property is ignored by the
Spark Dispatcher.
As a consequence, the Mesos slave where the Spark driver is running does not
comply to the given Mesos constraints, but on the other hand, the Mesos
constraints are well applied for the Spark executors (without the need to patch
anything).
*BUT* we do not want necessarily apply the same Mesos constraints for the
driver and executors.
For instance, we may need to run Spark drivers and executors on 2 exclusive
types of Mesos slaves:
- The dispatcher is given Mesos resources only for drivers
- Once a driver is launched, it becomes a Mesos framework itself and is
responsible for reserving resources for its executors
- If we schedule too many jobs on a Mesos cluster through the dispatcher, the
whole cluster is allocated for the drivers and there are no more resources
available for the executors. A driver may be launched but it may be waiting for
resources for its executors infinitely, which leads to a congestion then to a
dead-lock situation.
- A solution to work around this problem is to *not* mix the drivers and
executors on the same machines by passing different Mesos constraints for the
driver and for executors.
The 'spark.mesos.constraints' property still apply for executors. As for the
drivers, the 'spark.mesos.dispatcher.driverDefault.[PropertyName]' generic
property seems ideal. By definition, it allows to "set default properties for
drivers submitted through the dispatcher".
I propose to revise this patch and to use
'spark.mesos.dispatcher.driverDefault.spark.mesos.constraints' instead of
'spark.mesos.constraints'.
What do you guys think?
> Support constraints in spark-dispatcher
> ---------------------------------------
>
> Key: SPARK-19606
> URL: https://issues.apache.org/jira/browse/SPARK-19606
> Project: Spark
> Issue Type: New Feature
> Components: Mesos
> Affects Versions: 2.1.0
> Reporter: Philipp Hoffmann
>
> The `spark.mesos.constraints` configuration is ignored by the
> spark-dispatcher. The constraints need to be passed in the Framework
> information when registering with Mesos.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
