Reuse of Field Names in Haskell Records

The Problem: The current design of Haskell's record system does not allow for field names to be reused across different record types. For example, a programmer may not define two record types in the same module that both have a field called name. Even when the records are defined in different modules, it is inconvenient to work with records that have the same field name.

This page describes a point in the design space that lifts this restriction, and has the following additional benefits:

  • the change is simple and easy to describe to a Haskell programmer,
  • it is fully compatible with other record extensions, such as RecordWildCards and NamedFieldPuns,
  • it does not preclude further record system extensions for more sophisticated behavior,
  • the compiler needs to do less work than in the current record implementation.

The Design

The following paragraphs describe the various operations on records.


Records are declared in exactly the same way as they are in Haskell'98. However, we don't need to check that multiple record declarations in the same module have distinct field names.

For example:

data Person   = Person   { name :: String, age :: Int }
data Variable = Variable { name :: Int, varType :: Type }


Record values are constructed in exactly the same way as in Haskell'98. For example:

john = Person { name = "John", age = 10 }
var  = Variable { name = 0, varType = TInt }

There is no ambiguity when constructing values because the constructors are still unique for each datatype.


The update notation of Haskell'98 is not compatible with reusing fields across records, because it leads to ambiguity. For example, if we write \a -> x { name = a } we don't know if we are setting the field name for a Person, or Variable.

To work around this, we propose an alternative notation for record update, which is similar to the notation for record construction:

C { e | f1 = e1, f2 = e2 }

Here, C is a constructor, e is the record expression that is being updated, f1, f2, etc. are the fields that are being updates, and e1, e2, etc. are the new values of the fields.

For example, using the process package, we could write things like this:

main = do h <- openFile "stdout.txt" WriteMode
          createProcess CreateProcess { shell "ls" | std_out = h }

Unlike Haskell'98, this notation includes the constructor, so it is clear (both to programmers and the compiler) which field is being updated.

Also, as in Haskell'98, record update may change the types of fields. For example:

data Node a = Node { left :: a, node :: Int, right :: a }

makeBlank :: Node a -> Node ()
makeBlank n = Node { n | left = (), right = () }

Note that in this example it is important that we can update both fields at the same time---we could not change the fields one at a time, as the intermediate value would not be well typed. This type of update is supported by standard Haskell records, but it is not supported by most other proposals, including ones that generate lenses.

RAE: But this doesn't allow a record update to affect potentially more than one constructor. Currently, if multiple constructors of one type include the same field name, you can update all constructors. Given that this change is not backward compatible, what's the migration strategy? End RAE


We propose that no selector functions are automatically generated by the compiler!

While at first sight this might look a little drastic, the benefit is that the namespace is not polluted by automatically generated accessor functions. This frees programmers to access fields in whatever fashion is most convenient.

The simplest way to access a field, would be via pattern matching:

someFun1 C { x = v } = ...    -- traditional patterns
someFun2 C { x }     = ...    -- using punning
someFun3 C { .. }    = ...    -- using record wild cards

In addition, if there are no name clashes for some field, a programmer might choose to revert to the standard Haskell'98 behavior:

garfinkle :: MyType -> Int
garfinkle MyType { garfinkle = x } = x

On the other hand, if some other field appears in many types, and we need to access it often, then---as usual---we could define a class:

class Size a where
  size :: a -> Int

instance Size T1 where size T1 { size = x } = x
instance Size T2 where size T2 { size = x } = x
instance Size T3 where size T3 { length = x } = x

Of course, one could also manipulate fields through TH generated lenses, or whatever mechanism they find most convenient: we have a wide open space to experiment with various record accessing mechanisms.

Optional Extensions and Variations

Field Update With A Function

It is common to update the value of a field based on the previous value of the same field. In some cases, it might be cumbersome to pattern match just to get the old value for an update. In such situations, the following piece of syntactic sugar for record updates might be handy:

C { e | f1 = e1, f2 := e2 }

As before, C is a constructor, e is the record that is being updated, f1 is a field that is being set to the value of e1. The new syntax is in field f2: the := notation signifies that e2 is an updating function. This means that the new value for f2 will be obtained by applying e2 to the old value of f2 (i.e., the value that's in e).

nodeToBool :: Node Int -> Node Bool
nodeToBool n = Node { n | left := (< 0), right := (>= 0) }

increment :: Node a -> Node a
increment n = Node { n | middle := (+1) }

Variation of Update Notation

An alternative to the update notation would be to put the record that is being updated last:

C { f1 = e1, f2 = e2 | e }

Again C is the constructor, f1 and f2 are the fields being updated, and e is the record that is being updated.

Both notations seem compatible with parsing Haskell. This notation emphasizes the fields that are being changed, while the other emphasizes the old record. It is not immediately clear, which of the two variations is more readable in practice.

Last modified 5 years ago Last modified on Jan 29, 2015 12:43:56 PM