alexarchambault / data-class   0.2.6

Apache License 2.0 GitHub

case-classes with better binary compatibility story

Scala versions: 2.13 2.12 2.11

data-class

Build Status Maven Central

data-class allows to create classes almost like case-classes, but with no public unapply or copy methods, making it easier to add fields to them while maintaining binary compatiblity.

Usage

Setup

Add to your build.sbt,

libraryDependencies += "io.github.alexarchambault" %% "data-class" % "0.2.1"

The latest version is Maven Central.

The macro paradise plugin is needed up to scala 2.12, and the right compiler option needs to be used from 2.13 onwards:

lazy val isAtLeastScala213 = Def.setting {
  import Ordering.Implicits._
  CrossVersion.partialVersion(scalaVersion.value).exists(_ >= (2, 13))
}
libraryDependencies ++= {
  if (isAtLeastScala213.value) Nil
  else Seq(compilerPlugin("org.scalamacros" % "paradise" % "2.1.1" cross CrossVersion.full))
}
scalacOptions ++= {
  if (isAtLeastScala213.value) Seq("-Ymacro-annotations")
  else Nil
}

Lastly, if you know what you are doing, you can manage to have data-class be a compile-time only dependency.

API

Use a @data annotation instead of a case modifier, like

import dataclass.data

@data class Foo(n: Int, s: String)

This annotation adds a number of features, that can also be found in case classes:

  • sensible equals / hashCode / toString implementations,
  • apply methods in the companion object for easier creation,
  • extend the scala.Product trait (itself extending scala.Equal), and implement its methods,
  • extend the scala.Serializable trait.

It also adds things that differ from case classes:

  • add final modifier to the class,
  • for each field, add a corresponding with method (field count: Int generates a method withCount(count: Int) returning a new instance of the class with count updated).

Most notably, it does not generate copy or unapply methods, making binary compatibility much more tractable upon adding new fields (see below).

In the example above, the @data macro generates code like the following (modulo macro hygiene):

final class Foo(val n: Int, val s: String) extends Product with Serializable {

  def withN(n: Int) = new Foo(n = n, s = s)
  def withS(s: String) = new Foo(n = n, s = s)

  override def toString: String = {
    val b = new StringBuilder("Foo(")
    b.append(String.valueOf(n))
    b.append(", ")
    b.append(String.valueOf(s))
    b.append(")")
    b.toString
  }

  override def canEqual(obj: Any): Boolean = obj != null && obj.isInstanceOf[Foo]
  override def equals(obj: Any): Boolean = this.eq(obj.asInstanceOf[AnyRef]) || canEqual(obj) && {
    val other = obj.asInstanceOf[Foo]
    n == other.n && s == other.s
  })

  override def hashCode: Int = {
    var code = 17 + "Foo".##
    code = 37 * code + n.##
    code = 37 * code + s.##
    37 * code
  }

  private def tuple = (this.n, this.s)

  override def productArity: Int = 2
  override def productElement(n: Int): Any = n match {
    case 0 => this.n
    case 1 => this.s
    case n => throw new IndexOutOfBoundsException(n.toString)
  }
}

object Foo {
  def apply(n: Int, s: String): Foo = new Foo(n, s)
}

shapeless

By default, the classes annotated with @data now have a shape that shapeless.Generic handles:

import dataclass.data

@data class Foo(n: Int, d: Double)

import shapeless._
Generic[Foo] // works

Note that with shapeless 2.3.3 and prior versions, Generic derivation may fail if the body of the @data class contains vals or lazy vals, see shapeless issue #934.

Adding fields

In order to retain binary compatibility when adding fields, one should:

  • annotate the first added field with dataclass.since,
  • provide default values for the added fields, like
import dataclass._

@data class Foo(n: Int, d: Double, @since s: String = "", b: Boolean = false)

The @since annotation makes the @data macro generate apply methods compatible with those without the new fields.

The example above generates the following apply methods in the companion object of Foo:

object Foo {
  def apply(n: Int, d: Double): Foo = new Foo(n, d, "", false)
  def apply(n: Int, d: Double, s: String, b: Boolean) = new Foo(n, d, s, b)
}

The @since annotation accepts an optional string argument - a version can be passed for example - and it can be used multiple times, like

import dataclass._

@data class Foo(
  n: Int,
  d: Double,
  @since("1.1")
  s: String = "",
  b: Boolean = false,
  @since("1.2")
  count: Option[Int] = None,
  info: Option[String] = None
)

This generates the following apply methods in the companion object of Foo:

object Foo {
  def apply(n: Int, d: Double): Foo = new Foo(n, d, "", false, None, None)
  def apply(n: Int, d: Double, s: String, b: Boolean) = new Foo(n, d, s, b, None, None)
  def apply(n: Int, d: Double, s: String, b: Boolean, count: Option[Int], info: Option[String]) = new Foo(n, d, s, b, count, info)
}

Related work

  • contraband relies on code generation from JSON or a custom schema language to generate classes that can be evolved in a binary compatible way
  • stalagmite generates case classes with custom features via some macros (but doesn't aim at helping maintaining binary compatibility)