Scala PartialFunction Stackoverflow

Question

I'm developing a Scala / Python library called PySpark Cassandra. In it I have to deal with objects serialized Python objects in the pickle format when e.g. saving data.

I have a job which fails with a stackoverlfow:

org.apache.spark.SparkException: Job aborted due to stage failure: Task 250 in stage 2.0 failed 4 times, most recent failure: Lost task 250.3 in stage 2.0 (TID 411, sp-prod-adg02.priv.tgho.nl): java.lang.StackOverflowError
        at pyspark_cassandra.UnpickledUUIDConverter$$anonfun$convertPF$1.applyOrElse(Pickling.scala:121)
        at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:165)
        at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:166)
        at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:166)
        ...
        at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:166)
        at scala.PartialFunction$OrElse.applyOrElse(PartialFunction.scala:166)

The code which starts this trace is:

object UnpickledUUIDConverter extends TypeConverter[UUID] {
  val tt = typeTag[UUID]
  def targetTypeTag = tt
  def convertPF = { case holder: UUIDHolder => holder.uuid }
}

located at https://github.com/TargetHolding/pyspark-cassandra/blob/master/src/main/scala/pyspark_cassandra/Pickling.scala#L118 (for more details and context).

The UUIDHolder class is defined as:

class UUIDHolder {
  var uuid: UUID = null

  def __setstate__(values: HashMap[String, Object]): UUID = {
    val i = values.get("int").asInstanceOf[BigInteger]
    val buffer = ByteBuffer.wrap(i.toByteArray())
    uuid = new UUID(buffer.getLong(), buffer.getLong())
    uuid
  }
}

(the weird construct of this class is for compatibility with py4j and how UUID objects are pickled by Python)

But my understanding of Scala and the relationship between case blocks and PartialFunctions is rather limited. Especially how my case block relates to https://github.com/scala/scala/blob/2.10.x/src/library/scala/PartialFunction.scala#L166 (I'm running on Scala 2.10.5)

To worsen my situation :) I have difficulty reproducing the error consistently. It occurs in a Spark job on different nodes, but not al the time. I have a dataset for which the issue exists when saving that dataset. But I can't pin it to a particular record in the dataset.

I would not expect a StackOverflow in any case with this code. Any help would be very much appreciated!

score 1 · Accepted Answer · answered Mar 13 '16 at 20:50

1

To answer the simple bits:

Your case block is a partial function literal as explained here, aka a pattern-matching anonymous function. That works because the return type of convertPF is a partial function.

It gets an applyOrElse as described here, which avoids invoking ifDefined and then apply.

The OrElse on your stack is just what wraps pf1 orElse pf2. Its implementation of applyOrElse delegates to each PartialFunction.

A very long chain of pfi orElse pfi_++ could overflow the stack on evaluation, or orElse_i orElse (orElse_i++ orElse ...).

answered Mar 13 '16 at 20:50

som-snytt

39,429
2
47
129

Thanks @som-snytt for the answer. I was thinking along these lines. But I can't wrap my head around what would be the orElse in this case? I would expect that my partial function (the case block) is only defined for objects of type UUIDHolder and would raise a MatchError or something similar when invoked with something else (isDefinedAt would return false for that input). I don't really see where the recursion would come from. – Frens Jan Mar 14 '16 at 08:24
Just had a rubber duck session with a colleague which was helpful. It was indeed a long chain of orElse invocations. Caused by registering the UnpickledUUIDConverter many many times (in Spark Cassandra Connector) as the type converters were invoked using an orElse construct and the list of converters was to long it raises the StackOverflow. – Frens Jan Mar 14 '16 at 08:57

Scala PartialFunction Stackoverflow

1 Answers1