[
https://issues.apache.org/jira/browse/SPARK-22145?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Stavros Kontopoulos updated SPARK-22145:
----------------------------------------
Description:
There are two issues with driver re-starting on mesos using the supervise flag:
- We need to add spark.mesos.driver.frameworkId to the reloaded properties for
checkpointing, otherwise the new frameworkId propagated by the dispatcher will
be overwritten by the checkpointed data.
- Unique task ids are not used by the dispatcher:
https://issues.apache.org/jira/browse/MESOS-4737
https://issues.apache.org/jira/browse/MESOS-3070
This issue is the same in principle as in the case with standalone mode where
the master needs to re-launch drivers with a new appId (driverId) to deal with
net partitions.
was:
There are two issues with driver re-starting on mesos using the supervise flag:
- We need to add spark.mesos.driver.frameworkId to the reloaded properties for
checkpointing, otherwise the new frameworkId propagated by the dispatcher will
be overwritten by the checkpointed data.
- Unique task ids are not used by the dispatcher:
https://issues.apache.org/jira/browse/MESOS-4737
https://issues.apache.org/jira/browse/MESOS-3070
This issue is the same in principle as the case with standalone mode where the
master needs to re-launch drivers with a new appId (driverId) to deal with net
partitions, but it affects mesos not spark, since we use different driver ids
anyway when we re-launch a driver.
> Issues with driver re-starting on mesos (supervise)
> ---------------------------------------------------
>
> Key: SPARK-22145
> URL: https://issues.apache.org/jira/browse/SPARK-22145
> Project: Spark
> Issue Type: Bug
> Components: Mesos
> Affects Versions: 2.3.0
> Reporter: Stavros Kontopoulos
>
> There are two issues with driver re-starting on mesos using the supervise
> flag:
> - We need to add spark.mesos.driver.frameworkId to the reloaded properties
> for checkpointing, otherwise the new frameworkId propagated by the dispatcher
> will be overwritten by the checkpointed data.
> - Unique task ids are not used by the dispatcher:
> https://issues.apache.org/jira/browse/MESOS-4737
> https://issues.apache.org/jira/browse/MESOS-3070
> This issue is the same in principle as in the case with standalone mode where
> the master needs to re-launch drivers with a new appId (driverId) to deal
> with net partitions.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
