Skip to content

Latest commit

 

History

History
316 lines (218 loc) · 13.4 KB

CONTRIBUTING.md

File metadata and controls

316 lines (218 loc) · 13.4 KB

Developer Guide

This document describes the organization and coding conventions used in the library. It is intended for contributors, but may also be of interest to users.

Module Structure

scalaz-core should be kept lean and mean: type classes and instances, important data structures, and functions related to these.

scalaz.{iteratee, effect, concurrent} have each been moved a dedicated sub-project. Consider the same approach when adding large new features.

Type Classes

Inheritance

Type class extension is encoded with inheritance. For example, Monoid extends Semigroup.

Hierarchy

The type class hierarchy is configured in the build (GenTypeClass). The SBT command genTypeClasses will recreate all type classes, preserving chunks of code delimited by pairs of //// comments.

Do not add code, comments, or imports outside of these delimiters.

Methods

The abstract methods of a type class should be listed first. Derived methods follow under a comment // derived methods.

Method should not have symbolic identifiers. (Symbolic aliases are allowed under scalaz.syntax._.)

Methods parameters should be organized to support type inference when using the type class directly.

def map[A, B](fa: F[A])(f: A => B): F[B]  // good
def map[A, B](fa: F[A], f: A => B): F[B]  // bad

Use currying sparingly. Less indirection helps performance. Curried versions of methods may be offered as alternatives.

def fold[A, B](fa: F[A])(f: (A, B) => B): F[B] // good
def fold[A, B](fa: F[A])(f: A => B => B): F[B] // bad
def fold[A, B](fa: F[A]): A => B => B => F[B] // bad
final def foldCurried[A, B](fa: F[A]): (A => B => B) => F[B] = f => fold(fa)((a, b) => f(a)(b)) // okay

Add general purpose derived methods directly to the type class. These may be left unfinalized, to allow a child type class or a type class instance to override them for efficiency.

Data Structures

Minimize the number of top level declarations associated with a data structure. In particular, use the companion object to scope the elements of the ADT. This avoids clutter in Scaladoc and namespace pollution for users who import scalaz._.

Offer direct access to methods like map, flatMap, and foldMap directly as class methods in the data structure. Type class implementations should delegate to these. This makes these methods more discoverable via Scaladoc and autocompletion, and avoids unneeded indirection in cases when abstraction is not required.

Define type class instances in the same files as the data structure. Type class instance is described in more detail separately.

sealed abstract class MyDataStructure[A] {
   final def map[B](f: A => B): F[B] = ...
   final def flatMap[B](f: A => F[B]): F[B] = ...
}

object MyDataStructure extends MyDataStructureInstances with MyDataStructureFunctions {
   // Define type aliases in the companion directly; avoids fights with the compiler later.
   type MDS[A] = MyDataStructure[A]

   // apply method not suitable for mixing, define directly in the companion.
   def apply[A](a: A): MyDataStructure[A] = ...
}

// described separately.
sealed abstract class MyDataStructureInstances {
   import MyDataStructure._
   type MDS[X] = MyDataStructure

   implicit object mdsMonad extends Monad[MDS] {
      def point[A](a: => A) = MDS(a)

      def bind[A, B](fa: MDS[A])(f: A => MDS[B]) = fa flatMap f // delegate

      // override to use directly call map
      override def map[A, B](fa: MDS[A])(f: A => B) = fa map f // delegate
   }
}

// Design to allow the user to mix these functions into an object
trait MyDataStructureFunctions {
    final def emptyMds[A]: MyDataStructure[A] = ...
}

Type Class Instances

General

Type class instances are packaged in a class ClassifiedTypeInstances.

The implicit members within the Instances class should be named:

classifiedTypeInstance // the primary instance, e.g. implicit object optionInstance extends MonadPlus[Option] with Traverse[Option]. classifiedTypeTypeClassName // additional instances, e.g. implicit def optionSemigroup[A]

Scala and Java standard library

Instances are organized under scalaz.std according to the package of the classified type. The package prefix scala is omitted. Where a type is aliased under the package scala, the shorter path to the alias determines the location of the type class.

scala.Option => scalaz.std.option._ scala.List => scalaz.std.list._ scala.math.BigInt => scalaz.std.math.bigInt java.math.BigInteger => scalaz.std.java.math.bigInteger

In the file ClassifiedType.scala, define the class ClassifiedTypeInstances, and mix it into a) an object classifiedType, and b) scalaz.std.AllInstances.

In some cases, a group of related type class instances may be defined in a single file, such as for AnyVals or TupleN.

Scalaz Data Structures

Define the MyDatastructureInstances and mix this into object MyDataStructure.

Type Class Dependencies

A type class instances for F[X] may depend on an instance for X. This can be extended to dependencies on multiple instances.

For example:

(Semigroup[A], Semigroup[B]) => Semigroup[(A, B)] (Monoid[A], Monoid[B]) => Monoid[(A, B)]

This demands careful organization of these implicits to ensure that implicitly[Semigroup[(Int, Int)] resolves without ambiguity (both of these functions could provide that).

Here's how to organize the type class instances and implicits for Tuple2. See the comments inline.

sealed abstract class TupleInstances0 {
  // defined in a supertype of TupleInstances as a tie-breaker in case of ambiguity.
  // pass the type class instances into the type class implementation trait as members.
  // We need to choose different names for the parameters and the members to avoid shadowing.
  // Usually we suffix the parameter name with '0', in this case '_' is used for better readability.
  implicit def tuple2Semigroup[A1, A2](implicit A1_ : Semigroup[A1], A2_ : Semigroup[A2]): Semigroup[(A1, A2)] = new Tuple2Semigroup[A1, A2] {
    // WARNING: It's really easy to write `def A1 = A1` here!!
    implicit def A1 = A1_
    implicit def A2 = A2_
  }
}

sealed abstract class TupleInstances extends TupleInstances0 {
  // In a subtype of TupleInstances0. The order doesn't matter, but by convention we follow the subtyping
  // relationship of the type classes.
  implicit def tuple2Monoid[A1, A2](implicit A1_ : Monoid[A1], A2_ : Monoid[A2]): Monoid[(A1, A2)] = new Tuple2Monoid[A1, A2] {
    implicit def A1: Monoid[A1] = A1_
    implicit def A2: Monoid[A2] = A2_
  }

  // It would be okay to have `tuple2Bifunctor` in this trait, as there is no chance of ambiguity with
  // tuple2Monoid in an implicit search.
}

object tuple extends TupleInstances

//
// Type class implementation traits
//
private trait Tuple2Semigroup[A1, A2] extends Semigroup[(A1, A2)] {
  implicit def A1 : Semigroup[A1]  // Requirements about the type
  implicit def A2 : Semigroup[A2]

  def append(f1: (A1, A2), f2: => (A1, A2)): (A1, A2) = (
    A1.append(f1._1, f2._1),
    A2.append(f1._2, f2._2)
  )
}

// Inherits from `Tuple2Semigroup` to get `append`. Put `Tuple2Semigroup` at the end
// of the extends list to use its definitions preferentially. This becomes important
// when using `Monad` and `Traverse`.
//
// In some cases, you need to add `override` to the methods in these traits to allow this
// ordering of the extends list. For example, see `OptionTFunctor#map`
//
private trait Tuple2Monoid[A1, A2] extends Monoid[(A1, A2)] with Tuple2Semigroup[A1, A2] {
  implicit def A1 : Monoid[A1] // refining the implicit requirement
  implicit def A2 : Monoid[A2]

  def zero: (A1, A2) = (A1.zero, A2.zero)
}

This pattern is applied for type class instances for the transformers (e.g OptionT, EitherT), for instances over compositions and products, and for instances derived via an isomorphism.

Syntax

The methods of a type class that should be copied to the corresponding TypeClassOps object.

For example:

final class FunctorOps[F[_],A] private[syntax](val self: F[A])(implicit val F: Functor[F]) extends Ops[F[A]] {

  final def map[B](f: A => B): F[B] = F.map(self)(f)

  // such esoterica should be restricted to these wrappers.
  final def ∘[B](f: A => B): F[B] = F.map(self)(f)
}

Do not add non-trivial methods implementations to these wrappers. Instead, define the method on the type class and delegate. A user should be able to opt-out of the syntax package without losing access to functionality.

The type class generation machinery maintains these files, again, just fill in the gaps. The generated implicit conversions use the Unapply and Unapply2 types to maximize type inference, see this in action in scalaz.example.UnapplyInference. The same technique may be used in the methods in the syntax wrapper, see TraverseOps#sequence.

Standard library types, and Scalaz data structures may also have a syntax defined. These files and implicit conversions are created by hand. For example, see scalaz.syntax.TreeOps and scalaz.syntax.std.OptionOps. Remember to add these to either ToDataOps or ToAllStdOps, and to update Syntaxes.

All the syntactic implicit conversions are eventually mixed into the object Scalaz. However, it is now possible, and recommended, to be more fine grained with imports.

import scalaz._, syntax.applicative._, std.option._  // fine grained imports: preferred
import scalaz._, Scalaz._                            // über import: convenient, but clutters namespace and slows compiler/IDE

Because all the conversions are packaged in traits, it is possible for users to create objects that aggregate a smaller set of the imports.

General

Type Parameters

In the absence of a context specific name for the type parameters, use these:

  • Use A, B, C, X for kind *
  • Use F, G for kind * -> *.

Declare type parameters requiring type constructors first in type parameter declarations. This is an arbitrary choice, but the consistency is worth it.

class Foo[F[_], A, B] // good
class Foo[A, F[_], B] // bad

Define type lambdas using the kind-projector plugin. Raw type-lambdas should be defined with either lower case or greek letters as the type parameters. This helps to distinguish them from the applied types. Type Lambdas should use the kind-projector syntax. Here's some inline syntax examples:

F[X, ?]     ===  ({type λ[α] = F[X, α]})#λ
F[X, ?, ?]  ===  ({type λ[α, β] = F[X, α, β]})#λ
F[X, ?]     ===  ({type l[a] = F[X, a]})#l
F[X, ?, ?]  ===  ({type l[a, b] = F[X, a, b]})#l

If using Lambda Type Function syntax (ie. when the kind-projector's inline syntax is insufficient), use greek letters for parameters.

λ[α => (α, α)]  ===  ({type λ[α] = (α, α)})#λ
λ[α[_] => F[α]  ===  ({type λ[α[_]] = F[α]})#λ

For more information see https://github.com/typelevel/kind-projector

Parameters

Similarly, use these as defaults for parameters:

  • Use a for a parameter of type A
  • Use fa for a parameter of type F[A]
  • Use F for an implicit parameter of type TypeClass[F].

Type annotations

  • Annotate the return type of all public methods
    • Exception: overriding methods, including in type classes
    • Exception: implicits defining type class instances. (TODO Are there still any corner cases in scalac where this will bite us?)

Using type classes

  • Do not use scalaz.syntax._ to implement scalaz._. Instead use the type classes directly. Every type class companion object has an apply method to obtain an instance: use Monad[M].bind instead of implicitly[Monad[M]].bind

Build Errors

  • Use of package objects has led to intermittent failures in incremental compilation, such as "package scalaz.syntax refers to nonexisting symbol." or a NPE. Use clean build as a workaround.

OSGi Support

All JARs contain OSGi metadata and are usable in an OSGi container without modification. The metadata is generated automatically by sbt-osgi, which delegates the heavy lifting to bnd.

All packages of each JAR are exported with the package version set to the version of Scalaz. All bundles import all used packages, with some exceptions for optional imports (unsatisfied optional imports will result in ClassNotFoundExceptions at runtime if dependent code paths are exercised).

To maintain the OSGi metadata, the OsgiKeys.* settings should be updated appropriately as new packages are created or existing packages are removed or renamed. Further, care should be taken to ensure each package is wholly contained by a single JAR.

If a new project is added, the exported packages must be defined by declaring an OsgiKeys.exportedPackages setting, typically via the osgiExport method:

osgiExport("scalaz.newproject")

How can I help?

  • Port some examples, or create new ones, to get a feel for the new organization.
  • Port/Write test cases
  • Port a missing data structure
  • Add type class instances (most Show, Equal, Ordering are missing).
  • Documentation
    • Class level documentation for each type class.
    • Brief method documentation welcome
    • core/show-doc in SBT will build and pop up the scaladoc.
  • Review code base for consistency problems
  • Review type class hierarchy