Skip to content

[SPARK-10582] If a new AM restarts, the total number of executors should be in initial state in driver side. #8737

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
KaiXinXiaoLei wants to merge 5 commits into from
Closed

[SPARK-10582] If a new AM restarts, the total number of executors should be in initial state in driver side. #8737

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
7b93a84
add function
Sep 14, 2015
d7ed6dc
using function
Sep 14, 2015
564725b
a space
Sep 14, 2015
209f4da
scala style, thanks
Sep 14, 2015
258f146
change
Sep 17, 2015
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions core/src/main/scala/org/apache/spark/ExecutorAllocationManager.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,6 +232,14 @@ private[spark] class ExecutorAllocationManager(
executor.scheduleAtFixedRate(scheduleTask, 0, intervalMillis, TimeUnit.MILLISECONDS)
}

/**
* Change the value of numExecutorsTarget.
Copy link
Contributor

@andrewor14 andrewor14 Oct 17, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is too vague...

Reset this manager to the initial starting state.
This must be called if the cluster manager is restarted.

*/
def reSetNumExecutorsTarget(): Unit = {
Copy link

@vidma vidma Sep 14, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

reset ?

Copy link
Contributor

@vanzin vanzin Sep 21, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

+1 reset

Copy link
Contributor

@andrewor14 andrewor14 Oct 17, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I would just call this reset() since it needs to do more than just setting the target

logDebug(s"Now reset the value of numExecutorsTarget.")
Copy link
Contributor

@vanzin vanzin Sep 21, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

indentation is off

numExecutorsTarget = conf.getInt("spark.dynamicAllocation.initialExecutors", minNumExecutors)
Copy link
Contributor

@vanzin vanzin Sep 21, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If I understand correctly what's going on, this will result in addExecutors to issue a request for the desired number of executors, if that does not match spark.dynamicAllocation.initialExecutors, right? That same method will also bring numExecutorsTarget up to whatever was the previous value once it runs.

Instead, wouldn't it be clearer to just explicitly send a message to the AM saying "this is the current state of the world, initialize yourself to deal with it"? Maybe as a reply to the RegisterClusterManager message.

Also, YarnAllocator holds a lot of state about existing containers, such as in allocatedHostToContainersMap and releasedContainers. Is that data re-created somehow when the new AM comes up? If not, what are the side-effects, if any, of not having that information around?

Copy link
Contributor

@andrewor14 andrewor14 Oct 17, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I actually think we should reset on the driver side as this patch has done, since we end up calling updateAndSync anyway

}

/**
* Stop the allocation manager.
*/
Expand Down
1 change: 1 addition & 0 deletions core/src/main/scala/org/apache/spark/scheduler/cluster/YarnSchedulerBackend.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,6 +91,7 @@ private[spark] abstract class YarnSchedulerBackend(
filterParams.foreach { case (k, v) => conf.set(s"spark.$filterName.param.$k", v) }
scheduler.sc.ui.foreach { ui => JettyUtils.addFilters(ui.getHandlers, conf) }
}
scheduler.sc.executorAllocationManager.foreach(_.reSetNumExecutorsTarget())
Copy link

@vidma vidma Sep 14, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I guess this should go into some other place. you've put it here into addWebUIFilter, just temporary?

Copy link
Member

@srowen srowen Sep 14, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes this whole change needs a better explanation. At a high leve, why does it occur and why is this the right fix? This doesn't look right

Copy link
Author

@KaiXinXiaoLei KaiXinXiaoLei Sep 14, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

My way is： if a new AM is started, the value of numExecutorsTarget is reset. So the total number of executor will be calculated again, and driver will send RequestExecutors to AM to start execuots.

Copy link
Contributor

@jerryshao jerryshao Sep 16, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think it is not good to add this line of change here. I think your purpose is to detect whether the AM is reconnected, right? There's are many other ways for you to detect it, adding here is a little weird.

Copy link
Contributor

@jerryshao jerryshao Sep 16, 2015

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think you could add in here (https://github.com/apache/spark/blob/master/core/src/main/scala/org/apache/spark/scheduler/cluster/YarnSchedulerBackend.scala#L170)

}

override def createDriverEndpoint(properties: Seq[(String, String)]): DriverEndpoint = {
Expand Down