Monday, July 13, 2020

Record constructors

records

This is a short post documenting various record-related idioms in the Haskell ecosystem. First-time package users can use this post to better understand record API idioms they encounter in the wild.

For package authors, I also include a brief recommendation near the end of the post explaining which idiom I personally prefer.

The example

I’ll use the following record type as the running example for this post:

module Example where

data Person = Person{ name :: String , admin :: Bool }

There are a few ways you can create a Person record if the package author exports the record constructors.

The simplest approach requires no extensions. You can initialize the value of every field in a single expression, like this:

example :: Person
example = Person{ name = "John Doe", admin = True }

Some record literals can get quite large, so the language provides two extensions which can help with record assembly.

First, you can use the NamedFieldPuns extension, to author a record like this:

{-# LANGUAGE NamedFieldPuns #-}

example :: Person
example = Person{ name, admin }
  where
    name = "John Doe"

    admin = True

This works because the NamedFieldPuns extension translates Person{ name, admin } to Person{ name = name, admin = admin }.

The RecordWildCards extension goes a step further and allows you to initialize a record literal without naming all of the fields (again), like this:

{-# LANGUAGE RecordWildCards #-}

example :: Person
example = Person{..}
  where
    name = "John Doe"

    admin = True

Vice versa, you can destructure a record literal in a few ways. For example, you can access record fields using accessor functions:

render :: Person -> String
render person = name person ++ suffix
  where
    suffix = if admin person then " - Admin" else ""

… or you can pattern match on a record literal:

render :: Person -> String
render Person{ name = name, admin = admin } = name ++ suffix
  where
    suffix = if admin then " - Admin" else ""

… or you can use the NamedFieldPuns extension (which also works in reverse):

render :: Person -> String
render Person{ name, admin } = name ++ suffix
  where
    suffix = if admin then " - Admin" else ""

… or you can use the RecordWildCards extension (which also works in reverse):

render :: Person -> String
render Person{..} = name ++ suffix
  where
    suffix = if admin then " - Admin" else ""

Also, once the RecordDotSyntax extension is available you can use ordinary dot syntax to access record fields:

render :: Person -> String
render person = person.name ++ suffix
  where
    suffix = if person.admin then " - Admin" else ""

Opaque record types

Some Haskell packages will elect to not export the record constructor. When they do so they will instead provide a function that initializes a record value with all required fields and defaults the remaining fields.

For example, suppose the name field were required for our Person type and the admin field were optional (defaulting to False). The API might look like this:

module Example (
      Person(name, admin)
    , makePerson
    ) where

data Person = Person{ name :: String, admin :: Bool }

makePerson :: String -> Person
makePerson name = Person{ name = name, admin = False }

Carefully note that the module exports the Person type and all of the fields, but not the Person constructor. So the only way that a user can create a Person record is to use the makePerson “smart constructor”. The typical idiom goes like this:

example :: Person
example = (makePerson "John Doe"){ admin = True }

In other words, the user is supposed to initialize required fields using the “smart constructor” and then set the remaining non-required fields using record syntax. This works because you can update a record type using exported fields even if the constructor is not exported.

The wai package is one of the more commonly used packages that observes this idiom. For example, the Request record is opaque but the accessors are still exported, so you can create a defaultRequest and then update that Request using record syntax:

example :: Request
example = defaultRequest{ requestMethod = "GET", isSecure = True }

… and you can still access fields using the exported accessor functions:

requestMethod example

This approach also works in conjunction with NamedFieldPuns for assembly (but not disassembly), so something like this valid:

example :: Request
example = defaultRequest{ requestMethod, isSecure }
  where
    requestMethod = "GET"

    isSecure = True

However, this approach does not work with the RecordWildCards language extension.

Some other packages go a step further and instead of exporting the accessors they export lenses for the accessor fields. For example, the amazonka-* family of packages does this, leading to record construction code like this:

example :: PutObject
example =
    putObject "my-example-bucket" "some-key" "some-body"
    & poContentLength .~ Just 9
    & poStorageClass  .~ ReducedRedundancy

… and you access fields using the lenses:

view poContentLength example

My recommendation

I believe that package authors should prefer to export record constructors instead of using smart constructors. Specifically, the smart constructor idiom requires too much specialized language knowledge to create a record, something that should be an introductory task for a functional programming language.

Package authors typically justify smart constructors to improve API stability since they permit adding new default-valued fields in a backwards compatible way. However, I personally do not weight such stability highly (both as a package author and a package user) because Haskell is a typed language and these changes are easy for reverse dependencies to accommodate with the aid of the type-checker.

I place a higher premium on improving the experience for new contributors so that Haskell projects can more easily take root within a polyglot engineering organization. Management tends to be less reluctant to accept Haskell projects within their organization if they feel that other teams can confidently contribute to the Haskell code.

Future directions

One long-term solution that could provide the best of both worlds is if the language had first-class support for default-valued fields. In other words, perhaps you could author a record type like this:

data Person = Person{ name :: String , admin :: Bool = False }

… and then you could safely omit default-valued fields when initializing a record. Of course, I haven’t fully thought through the implications of such a change.

6 comments:

  1. In all `render` examples except the first, I think the condition should be `admin`, not `admin person`.

    ReplyDelete
  2. Could it be that the where clauses should begin with `suffix = `?

    ReplyDelete
  3. s/This words because/This works because

    ReplyDelete