[SPARK-25103] CompletionIterator may delay GC of completed resources

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 2.0.1, 2.1.0, 2.2.0, 2.3.0
• Component/s: Spark Core
• Labels: bulk-closed
• Assignee: Unassigned
• Reporter: Eyal Farago
• Created: 13/Aug/18 16:03
• Updated: 08/Oct/19 05:41
• Resolved: 08/Oct/19 05:41

Description

while working on SPARK-22713 , I fund (and partially fixed) a scenario in which an iterator is already exhausted but still holds a reference to some resources that can be GCed at this point.

However, these resources can not be GCed because of this reference.

the specific fix applied in SPARK-22713 was to wrap the iterator with a CompletionIterator that cleans it when exhausted, thing is that it's quite easy to get this wrong by closing over local variables or this reference in the cleanup function itself.

I propose solving this by modifying CompletionIterator to discard references to the wrapped iterator and cleanup function once exhausted.

 

• a dive into the code showed that most CompletionIterators are eventually used by 

  org.apache.spark.scheduler.ShuffleMapTask#runTask

  which does:

writer.write(rdd.iterator(partition, context).asInstanceOf[Iterator[_ <: Product2[Any, Any]]])

looking at 

org.apache.spark.shuffle.ShuffleWriter#write

implementations, it seems all of them first exhaust the iterator and then perform some kind of post-processing: i.e. merging spills, sorting, writing partitions files and then concatenating them into a single file... bottom line the Iterator may actually be 'sitting' for some time after being exhausted.