Skip to content

[SPARK-34601][SHUFFLE] Add spark.shuffle.markFileLostOnExecutorLost to not delete shuffle file on executor lost event #31715

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
hiboyang wants to merge 3 commits into from
Closed

[SPARK-34601][SHUFFLE] Add spark.shuffle.markFileLostOnExecutorLost to not delete shuffle file on executor lost event #31715

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ddb6d8f
Add spark.shuffle.markFileLostOnExecutorLost to not delete shuffle fi…
hiboyang Mar 3, 2021
de2d2b5
Update comment for spark.shuffle.markFileLostOnExecutorLost
hiboyang Mar 3, 2021
1418e8f
Add unit test for spark.shuffle.markFileLostOnExecutorLost
hiboyang Mar 4, 2021
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
9 changes: 9 additions & 0 deletions core/src/main/scala/org/apache/spark/internal/config/package.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2117,4 +2117,13 @@ package object config {
// batch of block will be loaded in memory with memory mapping, which has higher overhead
// with small MB sized chunk of data.
.createWithDefaultString("3m")

private[spark] val MARK_FILE_LOST_ON_EXECUTOR_LOST =
ConfigBuilder("spark.shuffle.markFileLostOnExecutorLost")

@c21 c21 Mar 4, 2021

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

I am having similar concern with @viirya and @attilapiros . I think we should not make it as a user-facing config. If we would like introducing a config for this anyway, it'd better to start with internal() config, but not user-facing.

Though spark.shuffle.manager can be arbitrary class, but in practice, they are just a handful of implementation in one company's environment, and ideally the infra developers should control which implementation to use, instead of user to control this. Similarly for whether to mark shuffle file lost should be controlled by developers team but not users.

Just share some context, in FB, we (spark developer) control these kinds of behavior transparently and make this invisible to spark user. This also helps us to migrate to newer implementation easier without worrying about users setting wrong config. The mixed case (query uses customized shuffle service and default shuffle service) can happen quite a bit in production, as we have rate limit for traffic on customized service, and need fallback.

@hiboyang hiboyang Mar 4, 2021 •
edited
Loading

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

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

Hi @c21 , thanks for the comment, and known you from FB! We previously discussed with people from FB about Cosco, but not for this specific issue. Just curious, do you guys modify your internal Spark distribution to make executor lost event not trigger driver to delete map output?

@c21 c21 Mar 4, 2021

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

do you guys modify your internal Spark distribution to make executor lost event not trigger driver to delete map output?

Yeah I think so, I can do a double check with cosco folks as well tomorrow. btw nice to know you as well!

@hiboyang hiboyang Mar 4, 2021

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

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

Cool, that will be interested to know, thanks!

@attilapiros attilapiros Mar 5, 2021

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

@c21 May I ask what is the status of open sourcing Cosco?
I just think it might help us to make better decisions about the Shuffle plugin API of Spark if we could look into that one too.

@c21 c21 Mar 5, 2021

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

@attilapiros - unfortunately there's no specific ETA now, but the team is working on it.

.doc("Mark shuffle files as lost when an executor is lost. If you are using a remote " +
"shuffle service such that shuffle files are not stored on the executor, consider " +

@dongjoon-hyun dongjoon-hyun Mar 3, 2021

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

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

Well. In this context, remote shuffle service is not external shuffle service in Apache Spark because we are ignoring env.blockManager.externalShuffleServiceEnabled here. We have better be clear on this.

"setting this to false.")
.version("3.2.0")
.booleanConf
.createWithDefault(true)
}
3 changes: 2 additions & 1 deletion core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2034,7 +2034,8 @@ private[spark] class DAGScheduler(
// if the cluster manager explicitly tells us that the entire worker was lost, then
// we know to unregister shuffle output. (Note that "worker" specifically refers to the process
// from a Standalone cluster, where the shuffle service lives in the Worker.)
val fileLost = workerHost.isDefined || !env.blockManager.externalShuffleServiceEnabled
val fileLost = (workerHost.isDefined || !env.blockManager.externalShuffleServiceEnabled) &&
env.blockManager.markFileLostOnExecutorLost

@c21 c21 Mar 4, 2021

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

shouldn't we check this config to not take effect, if spark.shuffle.manager is sort (the default one)?

removeExecutorAndUnregisterOutputs(
execId = execId,
fileLost = fileLost,
Expand Down
3 changes: 3 additions & 0 deletions core/src/main/scala/org/apache/spark/storage/BlockManager.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -179,6 +179,9 @@ private[spark] class BlockManager(
// same as `conf.get(config.SHUFFLE_SERVICE_ENABLED)`
private[spark] val externalShuffleServiceEnabled: Boolean = externalBlockStoreClient.isDefined

private[spark] val markFileLostOnExecutorLost: Boolean = conf.get(
config.MARK_FILE_LOST_ON_EXECUTOR_LOST)

private val remoteReadNioBufferConversion =
conf.get(Network.NETWORK_REMOTE_READ_NIO_BUFFER_CONVERSION)

Expand Down