Simple, hassle-free, zero-boilerplate serialization in Scala/Java similar to Python's Pickle?

Question

Is there a simple, hassle-free approach to serialization in Scala/Java that's similar to Python's pickle? Pickle is a dead-simple solution that's reasonably efficient in space and time (i.e. not abysmal) but doesn't care about cross-language accessibility, versioning, etc. and allows for optional customization.

What I'm aware of:

• Java's built-in serialization is infamously slow ([1], [2]), bloated, and fragile. Also have to mark classes as Serializable---annoying when there are things that are clearly serializable but which don't have that annotation (e.g. not many Point2D authors mark these Serializable).
• Scala's BytePickle requires a bunch of boilerplate for every type you want to pickle, and even then it doesn't work with (cyclic) object graphs.
• jserial: Unmaintained and doesn't seem to be that much faster/smaller than the default Java serialization.
• kryo: Cannot (de-)serialize objects with no 0-arg ctors, which is a severe limitation. (Also you have to register every class you plan to serialize, or else you get significant slowdowns/bloat, but even so it's still faster than pickle.)
• protostuff: AFAICT, you have to register every class you intend to serialize in advance in a "schema."

Kryo and protostuff are the closest solutions I've found, but I'm wondering if there's anything else out there (or if there's some way to use these that I should be aware of). Please include usage examples! Ideally also include benchmarks.

Answer 1

I actually think you'd be best off with kryo (I'm not aware of alternatives that offer less schema defining other than non-binary protocols). You mention that pickle is not susceptible to the slowdowns and bloat that kryo gets without registering classes, but kryo is still faster and less bloated than pickle even without registering classes. See the following micro-benchmark (obviously take it with a grain of salt, but this is what I could do easily):

Python pickle

import pickle
import time
class Person:
    def __init__(self, name, age):
        self.name = name
        self.age = age
people = [Person("Alex", 20), Person("Barbara", 25), Person("Charles", 30), Person("David", 35), Person("Emily", 40)]
for i in xrange(10000):
    output = pickle.dumps(people, -1)
    if i == 0: print len(output)
start_time = time.time()
for i in xrange(10000):
    output = pickle.dumps(people, -1)
print time.time() - start_time    

Outputs 174 bytes and 1.18-1.23 seconds for me (Python 2.7.1 on 64-bit Linux)

Scala kryo

import com.esotericsoftware.kryo._
import java.io._
class Person(val name: String, val age: Int)
object MyApp extends App {
  val people = Array(new Person("Alex", 20), new Person("Barbara", 25), new Person("Charles", 30), new Person("David", 35), new Person("Emily", 40))
  val kryo = new Kryo
  kryo.setRegistrationOptional(true)
  val buffer = new ObjectBuffer(kryo)
  for (i <- 0 until 10000) {
    val output = new ByteArrayOutputStream
    buffer.writeObject(output, people)
    if (i == 0) println(output.size)
  }
  val startTime = System.nanoTime
  for (i <- 0 until 10000) {
    val output = new ByteArrayOutputStream
    buffer.writeObject(output, people)
  }
  println((System.nanoTime - startTime) / 1e9)
}

Outputs 68 bytes for me and 30-40ms (Kryo 1.04, Scala 2.9.1, Java 1.6.0.26 hotspot JVM on 64-bit Linux). For comparison, it outputs 51 bytes and 18-25ms if I register the classes.

Comparison

Kryo uses about 40% of the space and 3% of the time as Python pickle when not registering classes, and about 30% of the space and 2% of the time when registering classes. And you can always write a custom serializer when you want more control.