Tuesday, September 11, 2012

Full (de)serialization for Play Json using general purpose macros

I've released a new version of akmacros library for Scala 2.10. The release includes a new macro called 'factory'. Thanks to the new macro it is possible to construct a class with a public default constructor just passing a function to the generated factory of the class. The function receives symbol of an argument of the constructor and returns value for this argument. If function returns None, then default value for the argument is evaluated (provided that it is defined for the given argument). Here is the example:
    scala> import info.akshaal.clazz._
    import info.akshaal.clazz._

    scala> case class Record (name: String, twitter: Option[String] = None)
    defined class Record

    scala> val recordFactory = factory[Any, Record]('castValue)
    recordFactory: (Symbol => Option[Any]) => Record = <function1>

    scala> val args = Map('name -> "Evgeny")
    args: scala.collection.immutable.Map[Symbol,String] = Map('name -> Evgeny)

    scala> recordFactory(args.get)
    res0: Record = Record(Evgeny,None)

Writing less boilerplate code for Play Json

Just to demonstrate the way macro functions can simplify your code I've created a small project on github. The project includes Json library from the Play2.0 framework mostly as-is. The only changes I did are related to making it work on scala 2.10 (upgraded it to patched jerkson and so on). You might be interested in this file. It is the only file that I implemented in that project in order to add support for macros to the Play Json. Following the same pattern, you can easily use almost any other Json library like this.

Lets look what it gives you without any reflections.. Lets start with something simple:
    scala> import info.akshaal.json.play._
    import info.akshaal.json.play._

    scala> import play.api.libs.json._
    import play.api.libs.json._

    scala> case class Simple(str: String, num: Int = new Random().nextInt())
    defined class Simple

    scala> implicit val simpleJsFactory = factory[Simple]('fromJson)
    simpleJsFactory: (Symbol => Option[play.api.libs.json.JsValue]) => Simple = <function1>

    scala> implicit val simpleJsFields = allFields[Simple]('jsonate)
    simpleJsFields: List[info.akshaal.clazz.Field[Simple,play.api.libs.json.JsValue,None.type]] = List(Field(str,<function1>,None), Field(num,<function1>,None))
We can serialize and then deserialize:
    scala> val obj = new Simple("123", 5)
    obj: Simple = Simple(123,5)

    scala> val objJs = Json.toJson(obj)
    objJs: play.api.libs.json.JsValue = {"str":"123","num":5}

    scala> val obj2 = Json.fromJson[Simple](objJs)
    obj2: Simple = Simple(123,5)
That was easy. Now lets try to leverage default value feature of the case class:
    scala> val halfObjJs = Json.parse(""" {"str":"test"} """)
    halfObjJs: play.api.libs.json.JsValue = {"str":"test"}

    scala> val halfObj = Json.fromJson[Simple](halfObjJs)
    halfObj: Simple = Simple(test,-813663852)

    scala> val halfObj = Json.fromJson[Simple](halfObjJs)
    halfObj: Simple = Simple(test,-948806581)

    scala> val halfObj = Json.fromJson[Simple](halfObjJs)
    halfObj: Simple = Simple(test,428745442)
Note that we parsed halfObjJs three times. And 'num' was always different. That is because value for 'num' is missing in the halfObjJs and so the expression for default value was used (which is Random().nextInt, see defintion of Simple class). How about parametrized classes? Lets try:
    scala> case class Event[T](kind: String, payloads: T)
    defined class Event

    scala> implicit def eventJsFields[T: Writes] = allFields[Event[T]]('jsonate)
    eventJsFields: [T](implicit evidence$1: play.api.libs.json.Writes[T])List[info.akshaal.clazz.Field[Event[T],play.api.libs.json.JsValue,None.type]]

    scala> implicit def eventJsFactory[T: Reads] = factory[Event[T]]('fromJson)
    eventJsFactory: [T](implicit evidence$1: play.api.libs.json.Reads[T])(Symbol => Option[play.api.libs.json.JsValue]) => Event[T]
And test:
    scala> val event = Event(kind = "strange", payloads = obj)
    event: Event[Simple] = Event(strange,Simple(123,5))

    scala> val eventJs = Json.toJson(event)
    eventJs: play.api.libs.json.JsValue = {"kind":"strange","payloads":{"str":"123","num":5}}

    scala> val event3 = Json.fromJson[Event[Simple]](eventJs)
    event3: Event[Simple] = Event(strange,Simple(123,5))
In the real world it is unlikely that you will use completely unrestricted types (like T in the example above) in your case classes with json. It is more likely that there will be a (sealed) trait and some set of its subtypes. Lets see how it works. First, we will define some more case classes:
    sealed trait Message
    case class QuitMessage(msg: String) extends Message
    case class Heartbeat(id: Int) extends Message

    case class Messages(userId: Int, messages: List[Message] = List.empty)
Now, lets define how to serialize our new case classes:
    implicit def messageWrites = matchingWrites[Message] {
        case m: QuitMessage => quitMessageJsFields.toWrites.writes(m)
        case m: Heartbeat   => heartbeatJsFields.toWrites.writes(m)

    implicit def quitMessageJsFields = allFields[QuitMessage]('jsonate)
    implicit def heartbeatJsFields = allFields[Heartbeat]('jsonate)
    implicit def messagesJsFields = allFields[Messages]('jsonate)
And if deserialization is needed, lets define how to do it:
    implicit def quitMessageJsFactory = factory[QuitMessage]('fromJson)
    implicit def heartbeatJsFactory = factory[Heartbeat]('fromJson)
    implicit def messagesJsFactory = factory[Messages]('fromJson)

    implicit def messageReads: Reads[Message] =
            jsHas('msg) -> quitMessageJsFactory,
            jsHas('id) -> heartbeatJsFactory
Finally, lets try it:
    scala> val event5 =
              Event(kind = "t1",
                    payloads = Messages(userId = 4, messages = List(Heartbeat(5), QuitMessage("bye!"))))
    event5: Event[Messages] = Event(t1,Messages(4,List(Heartbeat(5), QuitMessage(bye!))))

    scala> val event5Js = Json.toJson(event5)
    event5Js: play.api.libs.json.JsValue = {"kind":"t1","payloads":{"userId":4,"messages":[{"id":5},{"msg":"bye!"}]}}
Pay attention, that Json for QuitMessage and Heartbeat classes has no type information or anything! It's just plain json with domain fields only. That's why messageReads has those jsHas occurances in its implementation, that is a little help for identifying which json object is what subtype of Message. Lets see that reading from json still works:
    scala> val event6 = Json.fromJson[Event[Messages]](event5Js)
    even6: Event[Messages] = Event(t1,Messages(4,List(Heartbeat(5), QuitMessage(bye!))))
Actually there is another way to do the same. You can inject an extra field into json object when serializing one of subtypes of an abstract type and use it as a guidance for reconstructing objects from json. It is really easy. Lets modify messageReads and messageWrites like this:
    implicit def messageWrites = matchingWrites[Message] {
        case m: QuitMessage => quitMessageJsFields.extra('type -> 'quit).toWrites.writes(m)
        case m: Heartbeat   => heartbeatJsFields.extra('type -> 'heart).toWrites.writes(m)

    implicit def messageReads: Reads[Message] =
            jsHas('type -> 'quit)  -> quitMessageJsFactory,
            jsHas('type -> 'heart) -> heartbeatJsFactory
Now we test the new implementation.
    scala> val messages = List(Heartbeat(1), QuitMessage("Hello"), Heartbeat(99))
    messages: List[Product with Serializable with Message] = List(Heartbeat(1), QuitMessage(Hello), Heartbeat(99))

    scala> val messagesJs = Json.toJson(messages)
    messagesJs: play.api.libs.json.JsValue = [{"type":"heart","id":1},{"type":"quit","msg":"Hello"},{"type":"heart","id":99}]

    scala> val messages2 = Json.fromJson[List[Message]](messagesJs)
    messages2: List[Message] = List(Heartbeat(1), QuitMessage(Hello), Heartbeat(99))
That's not all. What if there is an information in a case class you don't want to reveal in JSON? Consider the following case class. I will annotate fields that are safe to export by Ok annotation:
    class Ok extends annotation.StaticAnnotation

    case class User(@Ok login: String,
                    @Ok fullName: String,
                    @Ok messages: Int = 0,
                    passwordHash: Option[String] = None)
Now it's quite natural to define json fields like this:
    implicit val userJsFields = annotatedFields[User, Ok]('jsonate)
Let see it in action:
    |       val user = User(login = "akshaal",
    |                       fullName = "Evgeny Chukreev",
    |                       messages = 10,
    |                       passwordHash = Some("4d18758602c08243d7c08f8c9e4463b0"))
    user: User = User(akshaal,Evgeny Chukreev,10,Some(4d18758602c08243d7c08f8c9e4463b0))

    scala> val userJs = Json.toJson(user)
    userJs: play.api.libs.json.JsValue = {"login":"akshaal","fullName":"Evgeny Chukreev","messages":10}
It works.. But you can do more. Recall (if you looked at the implementation) that jsonate function was defined like this:
    def jsonate[T: Writes](t: T, args: Any): JsValue = Json.toJson(t)
The function is used to make a json value out of field's value. It is applied to each field. So you can define your our function that does post-processing (pre-processing?) and use it with a macro. The second argument (args) might be a parameter set given to annotation, this gives you even more power for writing complex json serialization easily.

And not only JSON. Using the same approach you can (de)serialize object from/to XML...

Pros of akmacros-json

  • Domain classes are separated from any notion of JSON
  • Full control over serialization/deserialization
  • Easy to use
  • Easy to extend or implement your own
  • No runtime reflections used

Cons of akmacros-json

  • Depends on Jerkson which is officially unavailable for Scala 2.10 (you need to build it from my fork (in order to build it you will need also this forked project))
  • Includes a copy of Play Json
  • Scala 2.10-M7 has many bugs related to implicits, value classes... so implementation as you might have noticed is not perfect in terms of performance (defs are used instead of vals)
I hope things will change really soon with the release of Scala 2.10.

About general purpose macros

This tiny macros addition is built on top of Play JSON (which is built on top of Jerkson (which is built on top of Jackson)) and akmacros (which doesn't have dependencies). Checkout 78 lines long implementation here. See https://github.com/akshaal/akmacros for a bit more information about using akmacros with sbt.

Wednesday, September 5, 2012

Easily implementing Json serialization for Play using macro. Function by symbol (lisp-like)l

I liked the idea of using fields macro in scala so much so I created a dedicated project for this macro to start reusing it in the different projects I did. Here it is.

Few words about implementation of this macro

The code has been reworked and now it is possible to transform field value using a supplied function. It means that Field and Fields types, and fields macro signatures are changed. In addition, there is a new convenient macro called allFields. I.e.:
    case class Field[-I <: AnyRef, +R, +A <: Product](name: String, get: I => R, args: A)
    type Fields[-I <: AnyRef, +R, +A <: Product] = List[Field[I, R, A]]
    def fields[Ann, I <: AnyRef, R, Args <: Product](apply: Symbol) = macro fieldsImpl[Ann, I, R, Args]
    def allFields[I <: AnyRef, R](apply: Symbol) = macro fieldsImpl[Any, I, R, None.type]
allFields is the new function that lists all public value members enclosed in the given type regardless of annotations. R in a Field(s) type stands for RETURN and represents type of a value returned by the field getter transformed using the supplied function. Value transformer function is passed into the macro using its symbol. You might wonder why aren't we just using something like f : T => R forSome { type T } ? That's because you can't pass the following function that way:
    def trans[T : TypeClass](x : T) : Int = ???
(which is just a sweet way of saving some keystrokes by not writing this:)
    def trans[T](x : T)(implicit tc : TypeClass[T]) : Int = ???
i.e. you can't pass function with TWO parameters (one of which is implicit parameter) where a ONE parameter function is expected. That is quite obvious but anyway.. So we use Symbol (in the spirit of Lisp). As you might guess by looking at the snippet below, it is expected that the symbol is constructed directly and not passed by reference:
        val applyFunName =
            apply.tree match {
                case Apply(_, List(Literal(Constant(s)))) => s.toString
                case _ =>
                        "fields macro is expected to be used with symbol literal like 'nothing or 'myFunction")
Almost everything in the macro implementation remains more-less same, except part that constructs expression for getting field value out of object. Now it looks like this:
                val applyFunTree = c.parse(applyFunName)
                val getFunArgTree = ValDef(Modifiers(), newTermName("x"), TypeTree(instanceT), EmptyTree)
                val getFunBodyTree =
                        List(Select(Ident(newTermName("x")), newTermName(name)),
getFunBodyTree illustrates what signature is really expected for the transformer function: in addition to field value, all arguments of the annotation are passed into the function (or None if no annotation used or annotation has no arguments). For example, you can't use Predef.identity function, instead, you should use valueIdentity which is (already) defined like this:
def valueIdentity[X] (value : X, annotationArgs : Any) : X = value
Having annotation arguments provided for the currently processing field gives you possibility for further customization of how the value is transformed. Now lets do an example.

Real-world example

Suppose you want to serialize your custom classes into JSON with no boilerplate code what so ever. That is how you can do it with this only (general-purpose) macro. Lets define some generic Writes typeclase provider:
    implicit def writesForFields[T <: AnyRef](implicit fields: clazz.Fields[T, JsValue, _]): Writes[T] = {
        new Writes[T] {
            def writes(t: T): JsValue = {
                JsObject(fields map {
                    (field: clazz.Field[T, JsValue, _]) =>
                        field.name -> field.get(t: T)
The function shown above implicitly creates Writes for any type T which has an implicit instance of type Fields[T, JsValue, _] (read it like "List of fields of class T along with function to get value of type JsValue for each field"). Now lets define the transformer function, it will be used for serialization of field values:
def asJsValue[T : Writes](v : T, annArgs : Any) : JsValue = Json.toJson(v)
That was the only code needed to bootstrap your mini-serialization framework. Now you can use it. Lets assume you have declarations:
    case class JquerySocketEvent[T](id: Int, data: T, `type`: String = "message", reply: Boolean = false)
    case class ChatMessage(user: String, text: String)
    implicit def jquerySocketEventJsFields[T: Writes] = clazz.allFields[JquerySocketEvent[T], JsValue]('asJsValue)
    implicit val chatMessageJsFields = clazz.allFields[ChatMessage, JsValue]('asJsValue)
That's it. ... some fun:
        val event = JquerySocketEvent(id = 1, data = ChatMessage("Fluttershy", "Yay!"))
        println (Json.toJson(event))
... prints:
That was easy enough. Feel free to use it, re-implement it or implement a more powerful stuff. Macros FTW!