Uploaded image for project: 'Flink'
  1. Flink
  2. FLINK-23388

Non-static Scala case classes cannot be serialised

    XMLWordPrintableJSON

Details

    • Bug
    • Status: Open
    • Major
    • Resolution: Unresolved
    • None
    • None
    • None

    Description

      Explanation of the issue

      ScalaCaseClassSerializer is not powerful enough to serialise all Scala case classes that can be serialised in normal JVM serialisation (and the standard KryoSerializer). This is because it treats all case classes as made up only of their listed members. In fact, it is valid Scala to have other data in a case class, and in particular, it is possible for a nested case class to depend on data in its parent. This might be ill advised, but it is valid Scala:

      class Outer {
          var x = 0
          case class Inner(y: Int) {
              def z = x
          }
      }
      
      val outer = new Outer()
      val inner = outer.Inner(1)
      outer.x = 2
      
      scala> inner.z
      res0: Int = 2
      

      As of Scala 2.11, the compiler flag -Yrepl-class-based is made available (and defaults to on in Scala 2.13). It changes the way the Scala REPL and similar tools encapsulates the user code written in the REPL, wrapping it in a serialisable class rather than an object (static class). The purpose of this is to aid serialisation of the whole thing for distributed REPL systems that need to perform computation on remote nodes. One such application is flink-scala-shell, and another is Apache Zeppelin. See below for an illustration of the flag's importance in applications like these.

      In the JVM, a class can be serialised if it implements Serializable, and either it is static or its parent can be serialised. In the latter case the parent is brought with it to allow it to be constructed in its context at the other end.

      The Flink ScalaCaseClassSerializer does not understand that case classes (which always implement Serializable) might be nested inside a serialisable outer class. This is exactly the scenario that occurs when defining a supposedly top-level case class in flink-scala-shell or Zeppelin, because -Yrepl-class-based causes it to be nested inside a serialisable outer class. The consequence is that Scala case classes defined in one of these applications cannot be serialised in Flink, making them practically unusable. As a core Scala language feature, this is a serious omission from these projects.

      Fixing ScalaCaseClassSerializer - no free lunch

      I attempted to fix the issue by redirecting case classes in Flink to the standard KryoSerializer rather than the ScalaCaseClassSerializer, and at first glance this appeared to work very well. I was even able to run code in Zeppelin that sent a user-defined case class to be processed in a Flink job using the batch environment, and it worked well.

      Unfortunately, it didn't work when I tried to do the same thing using the streaming table environment. The error as presented was a failure to cast KryoSerializer to TupleSerializerBase, the superclass serialiser for tuples, products, and ScalaCaseClassSerializer. The Flink Table API assumes that case classes will be assigned a serialiser capable of moving to and fro from a table representation. This is a strong assumption that no case class instance will ever carry data besides its primary contents.

      In the case we're most interested in (the REPL class wrapper one), most case classes will not actually depend on additional data, but it's difficult for Flink to know that. In general, I imagine we would like to support cases where the case class does depend on additional data, although clearly Flink would be unable to turn it into a record and then back again. Perhaps we could do all the other operations without issue, but then raise an error if any operation attempted the latter transformation if there were missing additional data?

      Illustration of the importance of -Yrepl-class-based

      As illustration of why the flag is important, consider the following REPL interaction:

      val x = 1
      case class A(y: Int) {
          def z = x
      }
      
      val a = A(2)
      compute_remotely(() => a.z)
      

      The instance a requires the context x to be bought with it in order to perform the computation a.z, but if the REPL is placing the user code statically, then a simple serialisation of a will not take x with it.

      However, with the -Yrepl-class-based flag active, A is now a nested class which depends on the outer serialisable class of global state. x is now automatically transferred as part of JVM serialisation and deserialisation, and the REPL doesn't need to jump through any additional hoops to make this work.

      Attachments

        Issue Links

          Activity

            People

              Unassigned Unassigned
              tobymiller Toby Miller
              Votes:
              1 Vote for this issue
              Watchers:
              6 Start watching this issue

              Dates

                Created:
                Updated: