[SPARK-28917] Jobs can hang because of race of RDD.dependencies

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.3.3, 2.4.3
• Fix Version/s: 2.4.5, 3.0.0
• Component/s: Scheduler, Spark Core
• Assignee: Imran Rashid
• Reporter: Imran Rashid
• Created: 29/Aug/19 16:22
• Updated: 17/May/20 17:48
• Resolved: 08/Oct/19 18:43

Description

RDD.dependencies stores the precomputed cache value, but it is not thread-safe. This can lead to a race where the value gets overwritten, but the DAGScheduler gets stuck in an inconsistent state. In particular, this can happen when there is a race between the DAGScheduler event loop, and another thread (eg. a user thread, if there is multi-threaded job submission).

First, a job is submitted by the user, which then computes the result Stage and its parents:

https://github.com/apache/spark/blob/24655583f1cb5dae2e80bb572604fb4a9761ec07/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L983

Which eventually makes a call to rdd.dependencies:

https://github.com/apache/spark/blob/24655583f1cb5dae2e80bb572604fb4a9761ec07/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L519

At the same time, the user could also touch rdd.dependencies in another thread, which could overwrite the stored value because of the race.

Then the DAGScheduler checks the dependencies again later on in the job submission, via getMissingParentStages

https://github.com/apache/spark/blob/24655583f1cb5dae2e80bb572604fb4a9761ec07/core/src/main/scala/org/apache/spark/scheduler/DAGScheduler.scala#L1025

Because it will find new dependencies, it will create entirely different stages. Now the job has some orphaned stages which will never run.

One symptoms of this are seeing disjoint sets of stages in the "Parents of final stage" and the "Missing parents" messages on job submission (however this is not required).

(EDIT: Seeing repeated msgs "Registering RDD X" actually is just fine, it is not a symptom of a problem at all. It just means the RDD is the input to multiple shuffles.)

[INFO] 2019-08-15 23:22:31,570 org.apache.spark.SparkContext logInfo - Starting job: count at XXX.scala:462
...
[INFO] 2019-08-15 23:22:31,573 org.apache.spark.scheduler.DAGScheduler logInfo - Registering RDD 14 (repartition at XXX.scala:421)
...
...
[INFO] 2019-08-15 23:22:31,582 org.apache.spark.scheduler.DAGScheduler logInfo - Got job 1 (count at XXX.scala:462) with 40 output partitions
[INFO] 2019-08-15 23:22:31,582 org.apache.spark.scheduler.DAGScheduler logInfo - Final stage: ResultStage 5 (count at XXX.scala:462)
[INFO] 2019-08-15 23:22:31,582 org.apache.spark.scheduler.DAGScheduler logInfo - Parents of final stage: List(ShuffleMapStage 4)
[INFO] 2019-08-15 23:22:31,599 org.apache.spark.scheduler.DAGScheduler logInfo - Registering RDD 14 (repartition at XXX.scala:421)
[INFO] 2019-08-15 23:22:31,599 org.apache.spark.scheduler.DAGScheduler logInfo - Missing parents: List(ShuffleMapStage 6)

Another symptom is only visible with DEBUG logs turned on for DAGScheduler – you will calls to submitStage(Stage X) multiple times, followed by a different set of missing stages. eg. here, we see stage 1 first is missing stage 0 as a dependency, and then later on its missing stage 23:

19/09/19 22:28:15 DEBUG scheduler.DAGScheduler: submitStage(ShuffleMapStage 1)
19/09/19 22:28:15 DEBUG scheduler.DAGScheduler: missing: List(ShuffleMapStage 0)
...
19/09/19 22:32:01 DEBUG scheduler.DAGScheduler: submitStage(ShuffleMapStage 1)
19/09/19 22:32:01 DEBUG scheduler.DAGScheduler: missing: List(ShuffleMapStage 23)

Note that there is a similar issue w/ rdd.partitions. In particular for some RDDs, partitions references dependencies (eg. CoGroupedRDD).

There is also an issue that rdd.storageLevel is read and cached in the scheduler, but it could be modified simultaneously by the user in another thread. But, I can't see a way it could effect the scheduler.

WORKAROUND:
(a) call rdd.dependencies while you know that RDD is only getting touched by one thread (eg. in the thread that created it, or before you submit multiple jobs touching that RDD from other threads). Then that value will get cached.
(b) don't submit jobs from multiple threads.

Issue Links

• links to: GitHub Pull Request #25951