[ 
https://issues.apache.org/jira/browse/SPARK-13216?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15137797#comment-15137797
 ] 

Hari Shreedharan commented on SPARK-13216:
------------------------------------------

I disagree that checkpointing is only for failed applications. For any of the 
receiver-based streaming applications, checkpoints are important to recover as 
yet unprocessed data.

If the application cannot be reloaded from a checkpoint - then the old data is 
pretty much gone. I know that checkpointing basically makes application and 
spark upgrades difficult or impossible, but there are configuration parameters 
that the users might want to change based on load requirements etc. I don't see 
a reason why we should not allow this, since it has nothing to do with starting 
the app from checkpoint or not - if we want the number of executors to change 
we should be able to. This is especially true when migrating from a non-dynamic 
allocation situation to a dynamic allocation situation.

> Spark streaming application not honoring --num-executors in restarting of an 
> application from a checkpoint
> ----------------------------------------------------------------------------------------------------------
>
>                 Key: SPARK-13216
>                 URL: https://issues.apache.org/jira/browse/SPARK-13216
>             Project: Spark
>          Issue Type: Bug
>          Components: Spark Submit, Streaming
>    Affects Versions: 1.5.0
>            Reporter: Neelesh Srinivas Salian
>            Priority: Minor
>              Labels: Streaming
>
> Scenario to help understand:
> 1) The Spark streaming job with 12 executors was initiated with checkpointing 
> enabled.
> 2) In version 1.3, the user was able to append the number of executors to 20 
> using --num-executors but was unable to do so in version 1.5.
> In 1.5, the spark application still runs with 13 executors (1 for driver and 
> 12 executors).
> There is a need to start from the checkpoint itself and not restart the 
> application to avoid the loss of information.
> 3) Checked the code in 1.3 and 1.5, which shows the command 
> ''--num-executors" has been deprecated.
> Any thoughts on this? Not sure if anyone hit this one specifically before.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org

Reply via email to