Skip to content

Latest commit

 

History

History
1427 lines (1119 loc) · 46.8 KB

Functor.md

File metadata and controls

1427 lines (1119 loc) · 46.8 KB

Functor and Friends

Programming, like mathematics, is about abstraction. We try to model parts of the real world, reusing recurring patterns by abstracting over them.

In this chapter, we will learn about several related interfaces, which are all about abstraction and therefore can be hard to understand at the beginning. Especially figuring out why they are useful and when to use them will take time and experience. This chapter therefore comes with tons of exercises, most of which can be solved with only a few short lines of code. Don't skip them. Come back to them several times until these things start feeling natural to you. You will then realize that their initial complexity has vanished.

module Tutorial.Functor

import Data.List1
import Data.String
import Data.Vect

%default total

Functor

What do type constructors like List, List1, Maybe, or IO have in common? First, all of them are of type Type -> Type. Second, they all put values of a given type in a certain context. With List, the context is non-determinism: We know there to be zero or more values, but we don't know the exact number until we start taking the list apart by pattern matching on it. Likewise for List1, though we know for sure that there is at least one value. For Maybe, we are still not sure about how many values there are, but the possibilities are much smaller: Zero or one. With IO, the context is a different one: Arbitrary side effects.

Although the type constructors discussed above are quite different in how they behave and when they are useful, there are certain operations that keep coming up when working with them. The first such operation is mapping a pure function over the data type, without affecting its underlying structure.

For instance, given a list of numbers, we'd like to multiply each number by two, without changing their order or removing any values:

multBy2List : Num a => List a -> List a
multBy2List []        = []
multBy2List (x :: xs) = 2 * x :: multBy2List xs

But we might just as well convert every string in a list of strings to upper case characters:

toUpperList : List String -> List String
toUpperList []        = []
toUpperList (x :: xs) = toUpper x :: toUpperList xs

Sometimes, the type of the stored value changes. In the next example, we calculate the lengths of the strings stored in a list:

toLengthList : List String -> List Nat
toLengthList []        = []
toLengthList (x :: xs) = length x :: toLengthList xs

I'd like you to appreciate, just how boring these functions are. They are almost identical, with the only interesting part being the function we apply to each element. Surely, there must be a pattern to abstract over:

mapList : (a -> b) -> List a -> List b
mapList f []        = []
mapList f (x :: xs) = f x :: mapList f xs

This is often the first step of abstraction in functional programming: Write a (possibly generic) higher-order function. We can now concisely implement all examples shown above in terms of mapList:

multBy2List' : Num a => List a -> List a
multBy2List' = mapList (2 *)

toUpperList' : List String -> List String
toUpperList' = mapList toUpper

toLengthList' : List String -> List Nat
toLengthList' = mapList length

But surely we'd like to do the same kind of thing with List1 and Maybe! After all, they are just container types like List, the only difference being some detail about the number of values they can or can't hold:

mapMaybe : (a -> b) -> Maybe a -> Maybe b
mapMaybe f Nothing  = Nothing
mapMaybe f (Just v) = Just (f v)

Even with IO, we'd like to be able to map pure functions over effectful computations. The implementation is a bit more involved, due to the nested layers of data constructors, but if in doubt, the types will surely guide us. Note, however, that IO is not publicly exported, so its data constructor is unavailable to us. We can use functions toPrim and fromPrim, however, for converting IO from and to PrimIO, which we can freely dissect:

mapIO : (a -> b) -> IO a -> IO b
mapIO f io = fromPrim $ mapPrimIO (toPrim io)
  where mapPrimIO : PrimIO a -> PrimIO b
        mapPrimIO prim w =
          let MkIORes va w2 = prim w
           in MkIORes (f va) w2

From the concept of mapping a pure function over values in a context follow some derived functions, which are often useful. Here are some of them for IO:

mapConstIO : b -> IO a -> IO b
mapConstIO = mapIO . const

forgetIO : IO a -> IO ()
forgetIO = mapConstIO ()

Of course, we'd want to implement mapConst and forget as well for List, List1, and Maybe (and dozens of other type constructors with some kind of mapping function), and they'd all look the same and be equally boring.

When we come upon a recurring class of functions with several useful derived functions, we should consider defining an interface. But how should we go about this here? When you look at the types of mapList, mapMaybe, and mapIO, you'll see that it's the List, List1, and IO types we need to get rid of. These are not of type Type but of type Type -> Type. Luckily, there is nothing preventing us from parametrizing an interface over something else than a Type.

The interface we are looking for is called Functor. Here is its definition and an example implementation (I appended a tick at the end of the names for them not to overlap with the interface and functions exported by the Prelude):

interface Functor' (0 f : Type -> Type) where
  map' : (a -> b) -> f a -> f b

implementation Functor' Maybe where
  map' _ Nothing  = Nothing
  map' f (Just v) = Just $ f v

Note, that we had to give the type of parameter f explicitly, and in that case it needs to be annotated with quantity zero if you want it to be erased at runtime (which you almost always want).

Now, reading type signatures consisting only of type parameters like the one of map' can take some time to get used to, especially when some type parameters are applied to other parameters as in f a. It can be very helpful to inspect these signatures together with all implicit arguments at the REPL (I formatted the output to make it more readable):

Tutorial.Functor> :ti map'
Tutorial.Functor.map' :  {0 b : Type}
                      -> {0 a : Type}
                      -> {0 f : Type -> Type}
                      -> Functor' f
                      => (a -> b)
                      -> f a
                      -> f b

It can also be helpful to replace type parameter f with a concrete value of the same type:

Tutorial.Functor> :t map' {f = Maybe}
map' : (?a -> ?b) -> Maybe ?a -> Maybe ?b

Remember, being able to interpret type signatures is paramount to understanding what's going on in an Idris declaration. You must practice this and make use of the tools and utilities given to you.

Derived Functions

There are several functions and operators directly derivable from interface Functor. Eventually, you should know and remember all of them as they are highly useful. Here they are together with their types:

Tutorial.Functor> :t (<$>)
Prelude.<$> : Functor f => (a -> b) -> f a -> f b

Tutorial.Functor> :t (<&>)
Prelude.<&> : Functor f => f a -> (a -> b) -> f b

Tutorial.Functor> :t ($>)
Prelude.$> : Functor f => f a -> b -> f b

Tutorial.Functor> :t (<$)
Prelude.<$ : Functor f => b -> f a -> f b

Tutorial.Functor> :t ignore
Prelude.ignore : Functor f => f a -> f ()

(<$>) is an operator alias for map and allows you to sometimes drop some parentheses. For instance:

tailShowReversNoOp : Show a => List1 a -> List String
tailShowReversNoOp xs = map (reverse . show) (tail xs)

tailShowReverse : Show a => List1 a -> List String
tailShowReverse xs = reverse . show <$> tail xs

(<&>) is an alias for (<$>) with the arguments flipped. The other three (ignore, ($>), and (<$)) are all used to replace the values in a context with a constant. They are often useful when you don't care about the values themselves but want to keep the underlying structure.

Functors with more than one Type Parameter

The type constructors we looked at so far were all of type Type -> Type. However, we can also implement Functor for other type constructors. The only prerequisite is that the type parameter we'd like to change with function map must be the last in the argument list. For instance, here is the Functor implementation for Either e (note, that Either e has of course type Type -> Type as required):

implementation Functor' (Either e) where
  map' _ (Left ve)  = Left ve
  map' f (Right va) = Right $ f va

Here is another example, this time for a type constructor of type Bool -> Type -> Type (you might remember this from the exercises in the last chapter):

data List01 : (nonEmpty : Bool) -> Type -> Type where
  Nil  : List01 False a
  (::) : a -> List01 False a -> List01 ne a

implementation Functor (List01 ne) where
  map _ []        = []
  map f (x :: xs) = f x :: map f xs

Functor Composition

The nice thing about functors is how they can be paired and nested with other functors and the results are functors again:

record Product (f,g : Type -> Type) (a : Type) where
  constructor MkProduct
  fst : f a
  snd : g a

implementation Functor f => Functor g => Functor (Product f g) where
  map f (MkProduct l r) = MkProduct (map f l) (map f r)

The above allows us to conveniently map over a pair of functors. Note, however, that Idris needs some help with inferring the types involved:

toPair : Product f g a -> (f a, g a)
toPair (MkProduct fst snd) = (fst, snd)

fromPair : (f a, g a) -> Product f g a
fromPair (x,y) = MkProduct x y

productExample :  Show a
               => (Either e a, List a)
               -> (Either e String, List String)
productExample = toPair . map show . fromPair {f = Either e, g = List}

More often, we'd like to map over several layers of nested functors at once. Here's how to do this with an example:

record Comp (f,g : Type -> Type) (a : Type) where
  constructor MkComp
  unComp  : f (g a)

implementation Functor f => Functor g => Functor (Comp f g) where
  map f (MkComp v) = MkComp $ map f <$> v

compExample :  Show a => List (Either e a) -> List (Either e String)
compExample = unComp . map show . MkComp {f = List, g = Either e}

Named Implementations

Sometimes, there are more ways to implement an interface for a given type. For instance, for numeric types we can have a Monoid representing addition and one representing multiplication. Likewise, for nested functors, map can be interpreted as a mapping over only the first layer of values, or a mapping over several layers of values.

One way to go about this is to define single-field wrappers as shown with data type Comp above. However, Idris also allows us to define additional interface implementations, which must then be given a name. For instance:

[Compose'] Functor f => Functor g => Functor (f . g) where
  map f = (map . map) f

Note, that this defines a new implementation of Functor, which will not be considered during implicit resolution in order to avoid ambiguities. However, it is possible to explicitly choose to use this implementation by passing it as an explicit argument to map, prefixed with an @:

compExample2 :  Show a => List (Either e a) -> List (Either e String)
compExample2 = map @{Compose} show

In the example above, we used Compose instead of Compose', since the former is already exported by the Prelude.

Functor Laws

Implementations of Functor are supposed to adhere to certain laws, just like implementations of Eq or Ord. Again, these laws are not verified by Idris, although it would be possible (and often cumbersome) to do so.

  1. map id = id: Mapping the identity function over a functor must not have any visible effect such as changing a container's structure or affecting the side effects perfomed when running an IO action.

  2. map (f . g) = map f . map g: Sequencing two mappings must be identical to a single mapping using the composition of the two functions.

Both of these laws request, that map is preserving the structure of values. This is easier to understand with container types like List, Maybe, or Either e, where map is not allowed to add or remove any wrapped value, nor - in case of List - change their order. With IO, this can best be described as map not performing additional side effects.

Exercises part 1

  1. Write your own implementations of Functor' for Maybe, List, List1, Vect n, Either e, and Pair a.

  2. Write a named implementation of Functor for pairs of functors (similar to the one implemented for Product).

  3. Implement Functor for data type Identity (which is available from Control.Monad.Identity in base):

    record Identity a where
      constructor Id
      value : a
  4. Here is a curious one: Implement Functor for Const e (which is also available from Control.Applicative.Const in base). You might be confused about the fact that the second type parameter has absolutely no relevance at runtime, as there is no value of that type. Such types are sometimes called phantom types. They can be quite useful for tagging values with additional typing information.

    Don't let the above confuse you: There is only one possible implementation. As usual, use holes and let the compiler guide you if you get lost.

    record Const (e,a : Type) where
      constructor MkConst
      value : e
  5. Here is a sum type for describing CRUD operations (Create, Read, Update, and Delete) in a data store:

    data Crud : (i : Type) -> (a : Type) -> Type where
      Create : (value : a) -> Crud i a
      Update : (id : i) -> (value : a) -> Crud i a
      Read   : (id : i) -> Crud i a
      Delete : (id : i) -> Crud i a

    Implement Functor for Crud i.

  6. Here is a sum type for describing responses from a data server:

    data Response : (e, i, a : Type) -> Type where
      Created : (id : i) -> (value : a) -> Response e i a
      Updated : (id : i) -> (value : a) -> Response e i a
      Found   : (values : List a) -> Response e i a
      Deleted : (id : i) -> Response e i a
      Error   : (err : e) -> Response e i a

    Implement Functor for Repsonse e i.

  7. Implement Functor for Validated e:

    data Validated : (e,a : Type) -> Type where
      Invalid : (err : e) -> Validated e a
      Valid   : (val : a) -> Validated e a

Applicative

While Functor allows us to map a pure, unary function over a value in a context, it doesn't allow us to combine n such values under an n-ary function.

For instance, consider the following functions:

liftMaybe2 : (a -> b -> c) -> Maybe a -> Maybe b -> Maybe c
liftMaybe2 f (Just va) (Just vb) = Just $ f va vb
liftMaybe2 _ _         _         = Nothing

liftVect2 : (a -> b -> c) -> Vect n a -> Vect n b -> Vect n c
liftVect2 _ []        []        = []
liftVect2 f (x :: xs) (y :: ys) = f x y :: liftVect2 f xs ys

liftIO2 : (a -> b -> c) -> IO a -> IO b -> IO c
liftIO2 f ioa iob = fromPrim $ go (toPrim ioa) (toPrim iob)
  where go : PrimIO a -> PrimIO b -> PrimIO c
        go pa pb w =
          let MkIORes va w2 = pa w
              MkIORes vb w3 = pb w2
           in MkIORes (f va vb) w3

This behavior is not covered by Functor, yet it is a very common thing to do. For instance, we might want to read two numbers from standard input (both operations might fail), calculating the product of the two. Here's the code:

multNumbers : Num a => Neg a => IO (Maybe a)
multNumbers = do
  s1 <- getLine
  s2 <- getLine
  pure $ liftMaybe2 (*) (parseInteger s1) (parseInteger s2)

And it won't stop here. We might just as well want to have liftMaybe3 for ternary functions and three Maybe arguments and so on, for arbitrary numbers of arguments.

But there is more: We'd also like to lift pure values into the context in question. With this, we could do the following:

liftMaybe3 : (a -> b -> c -> d) -> Maybe a -> Maybe b -> Maybe c -> Maybe d
liftMaybe3 f (Just va) (Just vb) (Just vc) = Just $ f va vb vc
liftMaybe3 _ _         _         _         = Nothing

pureMaybe : a -> Maybe a
pureMaybe = Just

multAdd100 : Num a => Neg a => String -> String -> Maybe a
multAdd100 s t = liftMaybe3 calc (parseInteger s) (parseInteger t) (pure 100)
  where calc : a -> a -> a -> a
        calc x y z = x * y + z

As you'll of course already know, I am now going to present a new interface to encapsulate this behavior. It's called Applicative. Here is its definition and an example implementation:

interface Functor' f => Applicative' f where
  app   : f (a -> b) -> f a -> f b
  pure' : a -> f a

implementation Applicative' Maybe where
  app (Just fun) (Just val) = Just $ fun val
  app _          _          = Nothing

  pure' = Just

Interface Applicative is of course already exported by the Prelude. There, function app is an operator sometimes called app or apply: (<*>).

You may wonder, how functions like liftMaybe2 or liftIO3 are related to operator apply. Let me demonstrate this:

liftA2 : Applicative f => (a -> b -> c) -> f a -> f b -> f c
liftA2 fun fa fb = pure fun <*> fa <*> fb

liftA3 : Applicative f => (a -> b -> c -> d) -> f a -> f b -> f c -> f d
liftA3 fun fa fb fc = pure fun <*> fa <*> fb <*> fc

It is really important for you to understand what's going on here, so let's break these down. If we specialize liftA2 to use Maybe for f, pure fun is of type Maybe (a -> b -> c). Likewise, pure fun <*> fa is of type Maybe (b -> c), as (<*>) will apply the value stored in fa to the function stored in pure fun (currying!).

You'll often see such chains of applications of apply, the number of applies corresponding to the arity of the function we lift. You'll sometimes also see the following, which allows us to drop the initial call to pure, and use the operator version of map instead:

liftA2' : Applicative f => (a -> b -> c) -> f a -> f b -> f c
liftA2' fun fa fb = fun <$> fa <*> fb

liftA3' : Applicative f => (a -> b -> c -> d) -> f a -> f b -> f c -> f d
liftA3' fun fa fb fc = fun <$> fa <*> fb <*> fc

So, interface Applicative allows us to lift values (and functions!) into computational contexts and apply them to values in the same contexts. Before we will see an extended example why this is useful, I'll quickly introduce some syntactic sugar for working with applicative functors.

Idiom Brackets

The programming style used for implementing liftA2' and liftA3' is also referred to as applicative style and is used a lot in Haskell for combining several effectful computations with a single pure function.

In Idris, there is an alternative to using such chains of operator applications: Idiom brackets. Here's another reimplementation of liftA2 and liftA3:

liftA2'' : Applicative f => (a -> b -> c) -> f a -> f b -> f c
liftA2'' fun fa fb = [| fun fa fb |]

liftA3'' : Applicative f => (a -> b -> c -> d) -> f a -> f b -> f c -> f d
liftA3'' fun fa fb fc = [| fun fa fb fc |]

The above implementations will be desugared to the one given for liftA2 and liftA3, again before disambiguating, type checking, and filling in of implicit values. Like with the bind operator, we can therefore write custom implementations for pure and (<*>), and Idris will use these if it can disambiguate between the overloaded function names.

Use Case: CSV Reader

In order to understand the power and versatility that comes with applicative functors, we will look at a slightly extended example. We are going to write some utilities for parsing and decoding content from CSV files. These are files where each line holds a list of values separated by commas (or some other delimiter). Typically, they are used to store tabular data, for instance from spread sheet applications. What we would like to do is convert lines in a CSV file and store the result in custom records, where each record field corresponds to a column in the table.

For instance, here is a simple example file, containing tabular user information from a web store: First name, last name, age (optional), email address, gender, and password.

Jon,Doe,42,jon@doe.ch,m,weijr332sdk
Jane,Doe,,jane@doe.ch,f,aa433sd112
Stefan,Hoeck,,nope@goaway.ch,m,password123

And here are the Idris data types necessary to hold this information at runtime. We use again custom string wrappers for increased type safety and because it will allow us to define for each data type what we consider to be valid input:

data Gender = Male | Female | Other

record Name where
  constructor MkName
  value : String

record Email where
  constructor MkEmail
  value : String

record Password where
  constructor MkPassword
  value : String

record User where
  constructor MkUser
  firstName : Name
  lastName  : Name
  age       : Maybe Nat
  email     : Email
  gender    : Gender
  password  : Password

We start by defining an interface for reading fields in a CSV file and writing implementations for the data types we'd like to read:

interface CSVField a where
  read : String -> Maybe a

Below are implementations for Gender and Bool. I decided to in these cases encode each value with a single lower case character:

CSVField Gender where
  read "m" = Just Male
  read "f" = Just Female
  read "o" = Just Other
  read _   = Nothing

CSVField Bool where
  read "t" = Just True
  read "f" = Just False
  read _   = Nothing

For numeric types, we can use the parsing functions from Data.String:

CSVField Nat where
  read = parsePositive

CSVField Integer where
  read = parseInteger

CSVField Double where
  read = parseDouble

For optional values, the stored type must itself come with an instance of CSVField. We can then treat the empty string "" as Nothing, while a non-empty string will be passed to the encapsulated type's field reader. (Remember that (<$>) is an alias for map.)

CSVField a => CSVField (Maybe a) where
  read "" = Just Nothing
  read s  = Just <$> read s

Finally, for our string wrappers, we need to decide what we consider to be valid values. For simplicity, I decided to limit the length of allowed strings and the set of valid characters.

readIf : (String -> Bool) -> (String -> a) -> String -> Maybe a
readIf p mk s = if p s then Just (mk s) else Nothing

isValidName : String -> Bool
isValidName s =
  let len = length s
   in 0 < len && len <= 100 && all isAlpha (unpack s)

CSVField Name where
  read = readIf isValidName MkName

isEmailChar : Char -> Bool
isEmailChar '.' = True
isEmailChar '@' = True
isEmailChar c   = isAlphaNum c

isValidEmail : String -> Bool
isValidEmail s =
  let len = length s
   in 0 < len && len <= 100 && all isEmailChar (unpack s)

CSVField Email where
  read = readIf isValidEmail MkEmail

isPasswordChar : Char -> Bool
isPasswordChar ' ' = True
isPasswordChar c   = not (isControl c) && not (isSpace c)

isValidPassword : String -> Bool
isValidPassword s =
  let len = length s
   in 8 < len && len <= 100 && all isPasswordChar (unpack s)

CSVField Password where
  read = readIf isValidPassword MkPassword

In a later chapter, we will learn about refinement types and how to store an erased proof of validity together with a validated value.

We can now start to decode whole lines in a CSV file. In order to do so, we first introduce a custom error type encapsulating how things can go wrong:

data CSVError : Type where
  FieldError           : (line, column : Nat) -> (str : String) -> CSVError
  UnexpectedEndOfInput : (line, column : Nat) -> CSVError
  ExpectedEndOfInput   : (line, column : Nat) -> CSVError

We can now use CSVField to read a single field at a given line and position in a CSV file, and return a FieldError in case of a failure.

readField : CSVField a => (line, column : Nat) -> String -> Either CSVError a
readField line col str =
  maybe (Left $ FieldError line col str) Right (read str)

If we know in advance the number of fields we need to read, we can try and convert a list of strings to a Vect of the given length. This facilitates reading record values of a known number of fields, as we get the correct number of string variables when pattern matching on the vector:

toVect : (n : Nat) -> (line, col : Nat) -> List a -> Either CSVError (Vect n a)
toVect 0     line _   []        = Right []
toVect 0     line col _         = Left (ExpectedEndOfInput line col)
toVect (S k) line col []        = Left (UnexpectedEndOfInput line col)
toVect (S k) line col (x :: xs) = (x ::) <$> toVect k line (S col) xs

Finally, we can implement function readUser to try and convert a single line in a CSV-file to a value of type User:

readUser' : (line : Nat) -> List String -> Either CSVError User
readUser' line ss = do
  [fn,ln,a,em,g,pw] <- toVect 6 line 0 ss
  [| MkUser (readField line 1 fn)
            (readField line 2 ln)
            (readField line 3 a)
            (readField line 4 em)
            (readField line 5 g)
            (readField line 6 pw) |]

readUser : (line : Nat) -> String -> Either CSVError User
readUser line = readUser' line . forget . split (',' ==)

Let's give this a go at the REPL:

Tutorial.Functor> readUser 1 "Joe,Foo,46,j@f.ch,m,pw1234567"
Right (MkUser (MkName "Joe") (MkName "Foo")
  (Just 46) (MkEmail "j@f.ch") Male (MkPassword "pw1234567"))
Tutorial.Functor> readUser 7 "Joe,Foo,46,j@f.ch,m,shortPW"
Left (FieldError 7 6 "shortPW")

Note, how in the implementation of readUser' we used an idiom bracket to map a function of six arguments (MkUser) over six values of type Either CSVError. This will automatically succeed, if and only if all of the parsings have succeeded. It would have been notoriously cumbersome resulting in much less readable code to implement readUser' with a succession of six nested pattern matches.

However, the idiom bracket above looks still quite repetitive. Surely, we can do better?

A Case for Heterogeneous Lists

It is time to learn about a family of types, which can be used as a generic representation for record types, and which will allow us to represent and read rows in heterogeneous tables with a minimal amount of code: Heterogeneous lists.

namespace HList
  public export
  data HList : (ts : List Type) -> Type where
    Nil  : HList Nil
    (::) : (v : t) -> (vs : HList ts) -> HList (t :: ts)

A heterogeneous list is a list type indexed over a list of types. This allows us to at each position store a value of the type at the same position in the list index. For instance, here is a variant, which stores three values of types Bool, Nat, and Maybe String (in that order):

hlist1 : HList [Bool, Nat, Maybe String]
hlist1 = [True, 12, Nothing]

You could argue that heterogeneous lists are just tuples storing values of the given types. That's right, of course, however, as you'll learn the hard way in the exercises, we can use the list index to perform compile-time computations on HList, for instance when concatenating two such lists to keep track of the types stored in the result at the same time.

But first, we'll make use of HList as a means to concisely parse CSV-lines. In order to do that, we need to introduce a new interface for types corresponding to whole lines in a CSV-file:

interface CSVLine a where
  decodeAt : (line, col : Nat) -> List String -> Either CSVError a

We'll now write two implementations of CSVLine for HList: One for the Nil case, which will succeed if and only if the current list of strings is empty. The other for the cons case, which will try and read a single field from the head of the list and the remainder from its tail. We use again an idiom bracket to concatenate the results:

CSVLine (HList []) where
  decodeAt _ _ [] = Right Nil
  decodeAt l c _  = Left (ExpectedEndOfInput l c)

CSVField t => CSVLine (HList ts) => CSVLine (HList (t :: ts)) where
  decodeAt l c []        = Left (UnexpectedEndOfInput l c)
  decodeAt l c (s :: ss) = [| readField l c s :: decodeAt l (S c) ss |]

And that's it! All we need to add is two utility function for decoding whole lines before they have been split into tokens, one of which is specialized to HList and takes an erased list of types as argument to make it more convenient to use at the REPL:

decode : CSVLine a => (line : Nat) -> String -> Either CSVError a
decode line = decodeAt line 1 . forget . split (',' ==)

hdecode :  (0 ts : List Type)
        -> CSVLine (HList ts)
        => (line : Nat)
        -> String
        -> Either CSVError (HList ts)
hdecode _ = decode

It's time to reap the fruits of our labour and give this a go at the REPL:

Tutorial.Functor> hdecode [Bool,Nat,Double] 1 "f,100,12.123"
Right [False, 100, 12.123]
Tutorial.Functor> hdecode [Name,Name,Gender] 3 "Idris,,f"
Left (FieldError 3 2 "")

Applicative Laws

Again, Applicative implementations must follow certain laws. Here they are:

  • pure id <*> fa = fa: Lifting and applying the identity function has no visible effect.

  • [| f . g |] <*> v = f <*> (g <*> v): I must not matter, whether we compose our functions first and then apply them, or whether we apply our functions first and then compose them.

    The above might be hard to understand, so here they are again with explicit types and implementations:

    compL : Maybe (b -> c) -> Maybe (a -> b) -> Maybe a -> Maybe c
    compL f g v = [| f . g |] <*> v
    
    compR : Maybe (b -> c) -> Maybe (a -> b) -> Maybe a -> Maybe c
    compR f g v = f <*> (g <*> v)

    The second applicative law states, that the two implementations compL and compR should behave identically.

  • pure f <*> pure x = pure (f x). This is also called the homomorphism law. It should be pretty self-explaining.

  • f <*> pure v = pure ($ v) <*> f. This is called the law of interchange.

    This should again be explained with a concrete example:

    interL : Maybe (a -> b) -> a -> Maybe b
    interL f v = f <*> pure v
    
    interR : Maybe (a -> b) -> a -> Maybe b
    interR f v = pure ($ v) <*> f

    Note, that ($ v) has type (a -> b) -> b, so this is a function type being applied to f, which has a function of type a -> b wrapped in a Maybe context.

    The law of interchange states that it must not matter whether we apply a pure value from the left or right of the apply operator.

Exercises part 2

  1. Implement Applicative' for Either e and Identity.

  2. Implement Applicative' for Vect n. Note: In order to implement pure, the length must be known at runtime. This can be done by passing it as an unerased implicit to the interface implementation:

    implementation {n : _} -> Applicative' (Vect n) where
  3. Implement Applicative' for Pair e, with e having a Monoid constraint.

  4. Implement Applicative for Const e, with e having a Monoid constraint.

  5. Implement Applicative for Validated e, with e having a Semigroup constraint. This will allow us to use (<+>) to accumulate errors in case of two Invalid values in the implementation of apply.

  6. Add an additional data constructor of type CSVError -> CSVError -> CSVError to CSVError and use this to implement Semigroup for CSVError.

  7. Refactor our CSV-parsers and all related functions so that they return Validated instead of Either. This will only work, if you solved exercise 6.

    Two things to note: You will have to adjust very little of the existing code, as we can still use applicative syntax with Validated. Also, with this change, we enhanced our CSV-parsers with the ability of error accumulation. Here are some examples from a REPL session:

    Solutions.Functor> hdecode [Bool,Nat,Gender] 1 "t,12,f"
    Valid [True, 12, Female]
    Solutions.Functor> hdecode [Bool,Nat,Gender] 1 "o,-12,f"
    Invalid (App (FieldError 1 1 "o") (FieldError 1 2 "-12"))
    Solutions.Functor> hdecode [Bool,Nat,Gender] 1 "o,-12,foo"
    Invalid (App (FieldError 1 1 "o")
      (App (FieldError 1 2 "-12") (FieldError 1 3 "foo")))
    

    Behold the power of applicative functors and heterogeneous lists: With only a few lines of code we wrote a pure, type-safe, and total parser with error accumulation for lines in CSV-files, which is very convenient to use at the same time!

  8. Since we introduced heterogeneous lists in this chapter, it would be a pity not to experiment with them a little.

    This exercise is meant to sharpen your skills in type wizardry. It therefore comes with very few hints. Try to decide yourself what behavior you'd expect from a given function, how to express this in the types, and how to implement it afterwards. If your types are correct and precise enough, the implementations will almost come for free. Don't give up too early if you get stuck. Only if you truly run out of ideas should you have a glance at the solutions (and then, only at the types at first!)

    1. Implement head for HList.

    2. Implement tail for HList.

    3. Implement (++) for HList.

    4. Implement index for HList. This might be harder than the other three. Go back and look how we implemented indexList in an earlier exercise and start from there.

    5. Package contrib, which is part of the Idris project, provides Data.HVect.HVect, a data type for heterogeneous vectors. The only difference to our own HList is, that HVect is indexed over a vector of types instead of a list of types. This makes it easier to express certain operations at the type level.

      Write your own implementation of HVect together with functions head, tail, (++), and index.

    6. For a real challenge, try implementing a function for transposing a Vect m (HVect ts). You'll first have to be creative about how to even express this in the types.

      Note: In order to implement this, you'll need to pattern match on an erased argument in at least one case to help Idris with type inference. Pattern matching on erased arguments is forbidden (they are erased after all, so we can't inspect them at runtime), unless the structure of the value being matched on can be derived from another, un-erased argument.

      Also, don't worry if you get stuck on this one. It took me several tries to figure it out. But I enjoyed the experience, so I just had to include it here. :-)

      Note, however, that such a function might be useful when working with CSV-files, as it allows us to convert a table represented as rows (a vector of tuples) to one represented as columns (a tuple of vectors).

  9. Show, that the composition of two applicative functors is again an applicative functor by implementing Applicative for Comp f g.

  10. Show, that the product of two applicative functors is again an applicative functor by implementing Applicative for Prod f g.

Monad

Finally, Monad. A lot of ink has been spilled about this one. However, after what we already saw in the chapter about IO, there is not much left to discuss here. Monad extends Applicative and adds two new related functions: The bind operator ((>>=)) and function join. Here is its definition:

interface Applicative' m => Monad' m where
  bind  : m a -> (a -> m b) -> m b
  join' : m (m a) -> m a

Implementers of Monad are free to choose to either implement (>>=) or join or both. You will show in an exercise, how join can be implemented in terms of bind and vice versa.

The big difference between Monad and Applicative is, that the former allows a computation to depend on the result of an earlier computation. For instance, we could decide based on a string read from standard input whether to delete a file or play a song. The result of the first IO action (reading some user input) will affect, which IO action to run next. This is not possible with the apply operator:

(<*>) : IO (a -> b) -> IO a -> IO b

The two IO actions have already been decided on when they are being passed as arguments to (<*>). The result of the first cannot - in the general case - affect which computation to run in the second. (Actually, with IO this would theoretically be possible via side effects: The first action could write some command to a file or overwrite some mutable state, and the second action could read from that file or state, thus deciding on the next thing to do. But this is a speciality of IO, not of applicative functors in general. If the functor in question was Maybe, List, or Vector, no such thing would be possible.)

Let's demonstrate the difference with an example. Assume we'd like to enhance our CSV-reader with the ability to decode a line of tokens to a sum type. For instance, we'd like to decode CRUD requests from the lines of a CSV-file:

data Crud : (i : Type) -> (a : Type) -> Type where
  Create : (value : a) -> Crud i a
  Update : (id : i) -> (value : a) -> Crud i a
  Read   : (id : i) -> Crud i a
  Delete : (id : i) -> Crud i a

We need a way to on each line decide, which data constructor to choose for our decoding. One way to do this is to put the name of the data constructor (or some other tag of identification) in the first column of the CSV-file:

hlift : (a -> b) -> HList [a] -> b
hlift f [x] = f x

hlift2 : (a -> b -> c) -> HList [a,b] -> c
hlift2 f [x,y] = f x y

decodeCRUD :  CSVField i
           => CSVField a
           => (line : Nat)
           -> (s    : String)
           -> Either CSVError (Crud i a)
decodeCRUD l s =
  let h ::: t = split (',' ==) s
   in do
     MkName n <- readField l 1 h
     case n of
       "Create" => hlift  Create  <$> decodeAt l 2 t
       "Update" => hlift2 Update  <$> decodeAt l 2 t
       "Read"   => hlift  Read    <$> decodeAt l 2 t
       "Delete" => hlift  Delete  <$> decodeAt l 2 t
       _        => Left (FieldError l 1 n)

I added two utility function for helping with type inference and to get slightly nicer syntax. The important thing to note is, how we pattern match on the result of the first parsing function to decide on the data constructor and thus the next parsing function to use.

Here's how this works at the REPL:

Tutorial.Functor> decodeCRUD {i = Nat} {a = Email} 1 "Create,jon@doe.ch"
Right (Create (MkEmail "jon@doe.ch"))
Tutorial.Functor> decodeCRUD {i = Nat} {a = Email} 1 "Update,12,jane@doe.ch"
Right (Update 12 (MkEmail "jane@doe.ch"))
Tutorial.Functor> decodeCRUD {i = Nat} {a = Email} 1 "Delete,jon@doe.ch"
Left (FieldError 1 2 "jon@doe.ch")

To conclude, Monad, unlike Applicative, allows us to chain computations sequentially, where intermediary results can affect the behavior of later computations. So, if you have n unrelated effectful computations and want to combine them under a pure, n-ary function, Applicative will be sufficient. If, however, you want to decide based on the result of an effectful computation what computation to run next, you need a Monad.

Note, however, that Monad has one important drawback compared to Applicative: In general, monads don't compose. For instance, there is no Monad instance for Either e . IO. We will later learn about monad transformers, which can be composed with other monads.

Monad Laws

Without further ado, here are the laws for Monad:

  • ma >>= pure = ma and pure v >>= f = f v. These are monad's identity laws. Here they are as concrete examples:

    id1L : Maybe a -> Maybe a
    id1L ma = ma >>= pure
    
    id2L : a -> (a -> Maybe b) -> Maybe b
    id2L v f = pure v >>= f
    
    id2R : a -> (a -> Maybe b) -> Maybe b
    id2R v f = f v

    These two laws state that pure should behave neutrally w.r.t. bind.

  • (m >>= f) >>= g = m >>= (f >=> g). This is the law of associativity for monad. You might not have seen the second operator (>=>). It can be used to sequence effectful computations and has the following type:

    Tutorial.Functor> :t (>=>)
    Prelude.>=> : Monad m => (a -> m b) -> (b -> m c) -> a -> m c
    

The above are the official monad laws. However, we need to consider a third one, given that in Idris (and Haskell) Monad extends Applicative: As (<*>) can be implemented in terms of (>>=), the actual implementation of (<*>) must behave the same as the implementation in terms of (>>=):

  • mf <*> ma = mf >>= (\fun => map (fun $) ma).

Exercises part 3

  1. Applicative extends Functor, because every Applicative is also a Functor. Proof this by implementing map in terms of pure and (<*>).

  2. Monad extends Applicative, because every Monad is also an Applicative. Proof this by implementing (<*>) in terms of (>>=) and pure.

  3. Implement (>>=) in terms of join and other functions in the Monad hierarchy.

  4. Implement join in terms of (>>=) and other functions in the Monad hierarchy.

  5. There is no lawful Monad implementation for Validated e. Why?

  6. In this slightly extended exercise, we are going to simulate CRUD operations on a data store. We will use a mutable reference (imported from Data.IORef from the base library) holding a list of Users paired with a unique ID of type Nat as our user data base:

    DB : Type
    DB = IORef (List (Nat,User))

    Most operations on a database come with a risk of failure: When we try to update or delete a user, the entry in question might no longer be there. When we add a new user, a user with the given email address might already exist. Here is a custom error type to deal with this:

    data DBError : Type where
      UserExists        : Email -> Nat -> DBError
      UserNotFound      : Nat -> DBError
      SizeLimitExceeded : DBError

    In general, our functions will therefore have a type similar to the following:

    someDBProg : arg1 -> arg2 -> DB -> IO (Either DBError a)

    We'd like to abstract over this, by introducing a new wrapper type:

    record Prog a where
      constructor MkProg
      runProg : DB -> IO (Either DBError a)

    We are now ready to write us some utility functions. Make sure to follow the following business rules when implementing the functions below:

    • Email addresses in the DB must be unique. (Consider implementing Eq Email to verify this).

    • The size limit of 1000 entries must not be exceeded.

    • Operations trying to lookup a user by their ID must fail with UserNotFound in case no entry was found in the DB.

    You'll need the following functions from Data.IORef when working with mutable references: newIORef, readIORef, and writeIORef. In addition, functions Data.List.lookup and Data.List.find might be useful to implement some of the functions below.

    1. Implement interfaces Functor, Applicative, and Monad for Prog.

    2. Implement interface HasIO for Prog.

    3. Implement the following utility functions:

      throw : DBError -> Prog a
      
      getUsers : Prog (List (Nat,User))
      
      -- check the size limit!
      putUsers : List (Nat,User) -> Prog ()
      
      -- implement this in terms of `getUsers` and `putUsers`
      modifyDB : (List (Nat,User) -> List (Nat,User)) -> Prog ()
    4. Implement function lookupUser. This should fail with an appropriate error, if a user with the given ID cannot be found.

      lookupUser : (id : Nat) -> Prog User
    5. Implement function deleteUser. This should fail with an appropriate error, if a user with the given ID cannot be found. Make use of lookupUser in your implementation.

      deleteUser : (id : Nat) -> Prog ()
    6. Implement function addUser. This should fail, if a user with the given Email already exists, or if the data banks size limit of 1000 entries is exceeded. In addition, this should create and return a unique ID for the new user entry.

      addUser : (new : User) -> Prog Nat
    7. Implement function updateUser. This should fail, if the user in question cannot be found or a user with the updated user's Email already exists. The returned value should be the updated user.

      updateUser : (id : Nat) -> (mod : User -> User) -> Prog User
    8. Data type Prog is actually too specific. We could just as well abstract over the error type and the DB environment:

      record Prog' env err a where
        constructor MkProg'
        runProg' : env -> IO (Either err a)

      Verify, that all interface implementations you wrote for Prog can be used verbatim to implement the same interfaces for Prog' env err. The same goes for throw with only a slight adjustment in the function's type.

Background and further Reading

Concepts like functor and monad have their origin in category theory, a branch of mathematics. That is also where their laws come from. Category theory was found to have applications in programming language theory, especially functional programming. It is a highly abstract topic, but there is a pretty accessible introduction for programmers, written by Bartosz Milewski.

The usefulness of applicative functors as a middle ground between functor and monad was discovered several years after monads had already been in use in Haskell. They were introduced in the article Applicative Programming with Effects, which is freely available online and a highly recommended read.

Conclusion

  • Interfaces Functor, Applicative, and Monad abstract over programming patterns that come up when working with type constructors of type Type -> Type. Such data types are also referred to as values in a context, or effectful computations.

  • Functor allows us to map over values in a context without affecting the context's underlying structure.

  • Applicative allows us to apply n-ary functions to n effectful computations and to lift pure values into a context.

  • Monad allows us to chain effectful computations, where the intermediary results can affect, which computation to run further down the chain.

  • Unlike Monad, Functor and Applicative compose: The product and composition of two functors or applicatives are again functors or applicatives, respectively.

  • Idris provides syntactic sugar for working with some of the interfaces presented here: Idiom brackets for Applicative, do blocks and the bang operator for Monad.

What's next?

In the next chapter we get to learn more about recursion, totality checking, and an interface for collapsing container types: Foldable.