Structural types

November 3, 2018 • Scala types • Bartosz Konieczny

Home Scala types Structural types

Versions: Scala 2.12.1

Duck typing and Scala aren't the words going well together. However with a special kind of types we can achieve similar effect that in dynamically typed languages.

This post goes through the concept of structural types in Scala. Its first section shows the most basic points about it. The second one gives a little bit more examples of what we can do with these types. The last section has a micro-benchmark showing whether using structural type is penalizing or not.

Structural types defined

Since structural types are strongly related to duck typing, let's explain first this kind of "typing". Duck typing comes from dynamically typed languages and it's based on behavior rather than on static types:

function print_string(printable_instance):
  printable_instance.print()

As you can see, there is no expected type declaration at compile time. Thanks to that we can pass as parameter any object, as long as it implements print() method. In Scala stuctural type is a kind of specification that given object must met in order to be accepted as, for instance, input parameter. This specification has the form of { ...specification... } and it may look like:

describe("structural type") {
  it("should be used as method parameter") {
    object DotConcatenator {
      def concatenate(l1: String, l2: String): String = s"${l1}.${l2}"
    }
    object DashConcatenator{
      def concatenate(letter1: String, letter2: String): String = s"${letter1}-${letter2}"
    }

    def concatenateLetters(concatenateFunction: {def concatenate(letter1: String, letter2: String): String},
                            letters: Seq[String]): String = {
      letters.reduce((letter1, letter2) => concatenateFunction.concatenate(letter1, letter2))
    }

    val concatenatedWithDot = concatenateLetters(DotConcatenator, Seq("a", "b", "c"))
    val concatenatedWithDash = concatenateLetters(DashConcatenator, Seq("a", "b", "c"))

    concatenatedWithDot shouldEqual "a.b.c"
    concatenatedWithDash shouldEqual "a-b-c"
  }
}

As you can see, here we expect that the object provides a concatenate method taking 2 strings as parameters, regardless their names. We don't care about any other methods because they're not used in the implementation. Aside from methods, the structural type can also define fields:

it("should define values") {
  object SmallMultiplier {
    val MultiplyValue = 2
  }
  object BigMultiplier {
    val MultiplyValue = 4
  }

  def multiply(multiplier: {val MultiplyValue: Int}, nrToMultiply: Int): Int = {
    nrToMultiply * multiplier.MultiplyValue
  }

  val smallMultiplierResult = multiply(SmallMultiplier, 5)
  smallMultiplierResult shouldEqual 10
  val bigMultiplierResult = multiply(BigMultiplier, 5)
  bigMultiplierResult shouldEqual 20
}

You may wonder why should we need such construct in strongly typed languages as Scala ? Well, it brings some advantages. First, we don't need to create an enormous class hierarchy if we want to share some code among object behaving similarly. Thus, it brings flexibility by reducing codebase complexity. Also, despite lack of typing in the name, it's type-safe - Scala will check if passed type respects defined the specification. In the other side, it's much slower than regular methods because it uses reflection. The difference is especially visible for the case of loops and is shown in the last section.

Going deeper

The possibility to define method and fields in structural types isn't a single feature. We can also define more than one method or field:

it("should be used with 2 methods") {
  type MathOperator = {
    def add(nr1: Int, nr2: Int): Int
    def multiply(nr1: Int, nr2: Int): Int
  }
  def addAndMultiplyByTwo(mathOperator: MathOperator, nr1: Int, nr2: Int): Int = {
    mathOperator.multiply(mathOperator.add(nr1, nr2), 2)
  }

  val result = addAndMultiplyByTwo(new {
    def add(nr1: Int, nr2: Int) = nr1 + nr2
    def multiply(nr1: Int, nr2: Int) = nr1 * nr2
  }, 3, 5)

  result shouldEqual 16
}

In addition, structural type can be a part of first-order type:

it("should be used in first-order type") {
  object DotConcatenator {
    def concatenate(l1: String, l2: String): String = s"${l1}.${l2}"
  }
  object DashConcatenator{
    def concatenate(letter1: String, letter2: String): String = s"${letter1}-${letter2}"
  }

  class Concatenator[T <: {def concatenate(letter1: String, letter2: String): String}](concatenator: T) {
    def concatenateLetters(letter1: String, letter2: String): String = concatenator.concatenate(letter1, letter2)
  }

  val concatenatorsWithDots = new Concatenator(DotConcatenator)
  concatenatorsWithDots.concatenateLetters("a", "b") shouldEqual "a.b"

  val concatenatorWithDashes = new Concatenator(DashConcatenator)
  concatenatorWithDashes.concatenateLetters("a", "b") shouldEqual "a-b"
}

And if you bother about the readability problem by declaring structural types just with {}, you should know that they can be aliased:

it("should be created with new keyword from typed alias") {
  type Sum = {
    def add(nr1: Int, nr2: Int): Int
  }
  val sumImplementation = new {
    def add(nr1: Int, nr2: Int) = nr1 + nr2
  }
  def addNumbers(addFunc: Sum, nr1: Int, nr2: Int): Int = {
    addFunc.add(nr1, nr2)
  }

  val sumResult = addNumbers(sumImplementation, 1, 2)

  sumResult shouldEqual 3
}

Impact on loops

In order to estimate the impact of structural types on loops, I created a small micro-benchmark with JMH for the following code:

package com.waitingforcode.duck_typing

import java.util.concurrent.TimeUnit

import org.openjdk.jmh.annotations._

@OutputTimeUnit(TimeUnit.MILLISECONDS)
@BenchmarkMode(Array(Mode.All))
class DuckTypingMicroBenchmark {

  @Benchmark
  def verify_duck_typing: Unit = {
    type MathOperator = {
      def add(nr1: Int, nr2: Int): Int
      def multiply(nr1: Int, nr2: Int): Int
    }
    def addAndMultiplyBy2(mathOperator: MathOperator, nr1: Int, nr2: Int): Int = {
      mathOperator.multiply(mathOperator.add(nr1, nr2), 2)
    }
    for (nr <- 0 to 10000000) {
      addAndMultiplyBy2(MathOperatorAdding1ToEachValue, nr, nr+2)
    }
  }

  @Benchmark
  def verify_trait(): Unit = {
    def addAndMultiplyBy2(mathOperator: MathOperatorTrait, nr1: Int, nr2: Int): Int = {
      mathOperator.multiply(mathOperator.add(nr1, nr2), 2)
    }
    for (nr <- 0 to 10000000) {
      addAndMultiplyBy2(MathOperatorAdding1ToEachValueFromTrait, nr, nr+2)
    }
  }

}

trait MathOperatorTrait {
  def add(nr1: Int, nr2: Int): Int
  def multiply(nr1: Int, nr2: Int): Int
}

object MathOperatorAdding1ToEachValueFromTrait extends MathOperatorTrait {
  override def add(nr1: Int, nr2: Int): Int = nr1 + nr2 + 1 + 1

  override def multiply(nr1: Int, nr2: Int): Int = (nr1 + 1) * (nr2 + 1)
}

object MathOperatorAdding1ToEachValue {
  def add(nr1: Int, nr2: Int): Int = nr1 + nr2 + 1 + 1
  def multiply(nr1: Int, nr2: Int): Int = (nr1 + 1) * (nr2 + 1)
}

SBT files look like:

// build.sbt
import sbt.Keys._

lazy val root = (project in file(".")).
  enablePlugins(JmhPlugin).
  settings(
    name := "micro_benchmark",
    version := "1.0",
    scalaVersion := "2.12.6"
  )

// build.properties
sbt.version = 1.0.2

// plugins.sbt
logLevel := Level.Warn
addSbtPlugin("pl.project13.scala" % "sbt-jmh" % "0.3.2")

The microbenchmark ran with sbt jmh:run -i 20 -wi 10 -f1 -t1 -rf json command where: 20 is the number of measurement iterations, 10 is the number of warm-up iterations, f1 is the number of forks (benchmark ran in a subprocess) and t1 is the number of threads used in the benchmark. The last 2 options are related to the output format. After benchmarking the methods we can clearly see that structural types perform worse for all measurements:

Structural types are a convenient way to imitate duck-typing in strongly typed language as Scala. Unlike classical class hierarchy inheritance, structural types prefer to define a part of behavior that given object must provide. It reduces the complexity of hierarchies but in other side may lead to performance issues because of reflection used at resolution time. As proven in the 3rd part, when used with loops, structural types perform much worse than regular type-based constraints.

TAGS: #One Scala feature per week

Read also about Structural types here:

If you liked it, you should read:

privacy policy © 2014 - 2024 waitingforcode.com. All rights reserved | Design: Jakub Kędziora