Skip to content

[SPARK-17644] [CORE] Do not add failedStages when abortStage for fetch failure #15213

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
scwf wants to merge 9 commits into from
Closed

[SPARK-17644] [CORE] Do not add failedStages when abortStage for fetch failure #15213

Show file tree
Hide file tree
Changes from 7 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
8e667f5
test case
scwf Sep 23, 2016
2bfa05b
The failed stage never resubmitted
scwf Sep 23, 2016
d02cf93
test case
scwf Sep 23, 2016
7056cd6
fix style
scwf Sep 23, 2016
1f7bd88
test case imporvement
scwf Sep 23, 2016
d92adfc
comment fix
scwf Sep 23, 2016
1127ca1
address comments of zsxwing
scwf Sep 23, 2016
f91d86f
improve test
scwf Sep 28, 2016
09077cb
comment fix
scwf Sep 28, 2016
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: 7 additions & 2 deletions 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 @@ -1241,6 +1241,7 @@ class DAGScheduler(
s" ${task.stageAttemptId} and there is a more recent attempt for that stage " +
s"(attempt ID ${failedStage.latestInfo.attemptId}) running")
} else {
var abortedStage = false
// It is likely that we receive multiple FetchFailed for a single stage (because we have
// multiple tasks running concurrently on different executors). In that case, it is
// possible the fetch failure has already been handled by the scheduler.
Expand All @@ -1256,11 +1257,13 @@ class DAGScheduler(
if (disallowStageRetryForTest) {
abortStage(failedStage, "Fetch failure will not retry stage due to testing config",
None)
abortedStage = true
} else if (failedStage.failedOnFetchAndShouldAbort(task.stageAttemptId)) {
abortStage(failedStage, s"$failedStage (${failedStage.name}) " +
s"has failed the maximum allowable number of " +
s"times: ${Stage.MAX_CONSECUTIVE_FETCH_FAILURES}. " +
s"Most recent failure reason: ${failureMessage}", None)
abortedStage = true

@zsxwing zsxwing Sep 23, 2016

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.

There is another abortStage in if (disallowStageRetryForTest) { branch.

} else if (failedStages.isEmpty) {

@kayousterhout kayousterhout Sep 28, 2016

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.

Instead of having the abortedStage variable, how about re-writing the "else if" statement to be:

else {
if (failedStages.isEmpty) {
... stuff currently in else-if ...
}
failedStages += failedStage
failedStages += mapStage
}

That eliminates the confusion of multiple abortStage variables, as @zsxwing pointed out, and also makes the relationship between (i) adding the stage to failed stages and (ii) scheduling the Resubmit event more clear.

@scwf scwf Sep 28, 2016

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

it make sense to me, updated

// Don't schedule an event to resubmit failed stages if failed isn't empty, because
// in that case the event will already have been scheduled.
Expand All @@ -1271,8 +1274,10 @@ class DAGScheduler(
override def run(): Unit = eventProcessLoop.post(ResubmitFailedStages)
}, DAGScheduler.RESUBMIT_TIMEOUT, TimeUnit.MILLISECONDS)
}
failedStages += failedStage
failedStages += mapStage
if (!abortedStage) {
failedStages += failedStage
failedStages += mapStage
}
// Mark the map whose fetch failed as broken in the map stage
if (mapId != -1) {
mapStage.removeOutputLoc(mapId, bmAddress)
Expand Down
35 changes: 35 additions & 0 deletions core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,7 @@ import org.apache.spark._
import org.apache.spark.broadcast.BroadcastManager
import org.apache.spark.rdd.RDD
import org.apache.spark.scheduler.SchedulingMode.SchedulingMode
import org.apache.spark.shuffle.FetchFailedException

@kayousterhout kayousterhout Sep 28, 2016

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.

nit: can you group this with the next import (so import org.apache.spark.shuffle.{FetchFailedException, MetadataFetchFailedException}

import org.apache.spark.shuffle.MetadataFetchFailedException
import org.apache.spark.storage.{BlockId, BlockManagerId, BlockManagerMaster}
import org.apache.spark.util.{AccumulatorContext, AccumulatorV2, CallSite, LongAccumulator, Utils}
Expand Down Expand Up @@ -2105,6 +2106,40 @@ class DAGSchedulerSuite extends SparkFunSuite with LocalSparkContext with Timeou
assert(scheduler.getShuffleDependencies(rddE) === Set(shuffleDepA, shuffleDepC))
}

test("The failed stage never resubmitted due to abort stage in another thread") {

@squito squito Sep 27, 2016 •
edited
Loading

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.

A couple of changes for this test:

(a) Earlier there was already some discussion around naming and the PR title has been updated, this test should be renamed as well since multiple threads really have nothing to do with it.
(b) I'd prefer if tests are named to indicate the positive behavior that we want to verify . So with the above, I'd suggest a name like "After one stage is aborted for too many failed attempts, subsequent stages still behave correctly on fetch failures"
(c) duplicated code can be cleaned up (at first when I read the code, I was looking for differences between the two calls, so though its only one copy-paste, the intent is a lot clearer if its just once).

(d) I'd think it would be nice to also include a job which succeeds after a fetch failure at the end (3 jobs total). Unfortunately this is a bit of a pain to do in a test right now since you don't have access to stageAttemptId, but you can do it with something like this:

...
rdd1.map {
        case (x, _) if (x == 1) && FailThisAttempt._fail.getAndSet(false) =>
          throw new FetchFailedException(
            BlockManagerId("1", "1", 1), shuffleHandle.shuffleId, 0, 0, "test")
...

with helper

object FailThisAttempt extends Logging {
  val _fail = new AtomicBoolean(true)
}

failAfter(60.seconds) {
val e = intercept[SparkException] {
val rdd1 = sc.makeRDD(Array(1, 2, 3, 4), 2).map(x => (x, 1)).groupByKey()
val shuffleHandle =
rdd1.dependencies.head.asInstanceOf[ShuffleDependency[_, _, _]].shuffleHandle
rdd1.map {
case (x, _) if (x == 1) =>
throw new FetchFailedException(
BlockManagerId("1", "1", 1), shuffleHandle.shuffleId, 0, 0, "test")
case (x, _) => x
}.count()
}
assert(e.getMessage.contains("org.apache.spark.shuffle.FetchFailedException"))
}

// The following job that fails due to fetching failure will hang without
// the fix for SPARK-17644

@kayousterhout kayousterhout Sep 28, 2016

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.

Can you change to something like "Run a second job that will fail due to a fetch failure. This job will hang without the fix for SPARK-17644."

failAfter(60.seconds) {

@kayousterhout kayousterhout Sep 28, 2016

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 think a shorter timeout would be appropriate here to avoid slow-ness when this fails...maybe 10 seconds? That still seems plenty conservative since the resubmit timeout is 200 millis.

val e = intercept[SparkException] {
val rdd2 = sc.makeRDD(Array(1, 2, 3, 4), 2).map(x => (x, 1)).groupByKey()
val shuffleHandle =
rdd2.dependencies.head.asInstanceOf[ShuffleDependency[_, _, _]].shuffleHandle
rdd2.map {
case (x, _) if (x == 1) =>
throw new FetchFailedException(
BlockManagerId("1", "1", 1), shuffleHandle.shuffleId, 0, 0, "test")
case (x, _) => x
}.count()
}
assert(e.getMessage.contains("org.apache.spark.shuffle.FetchFailedException"))
}
}

/**
* Assert that the supplied TaskSet has exactly the given hosts as its preferred locations.
* Note that this checks only the host and not the executor ID.
Expand Down