Working with postgresql-simple with generics-sop

The least interesting part of my job as a programmer is the act of pressing keys on a keyboard, and thus I actively seek ways to reduce typing. As programmers, we aim for reuse in a our programs - abstracting commonality into reusable functions such that our programs get more concise. Functional programmers are aware of the benefits of higher-order functions as one form of generic programming, but another powerful technique is that of data type generic programming.

This variant of generic programming allows one to build programs that work over arbitrary data types, providing they have some sort of known “shape”. We describe the shape of data types by representing them via a code - often we can describe a data type as a sum of products. By sum, we are talking about the choice of a constructor in a data type (such as choosing between Left and Right to construct Either values), and by product we mean the individual fields in a constructor (such as the individual fields in a record).

Last month, Edsko and Löh announced a new library for generic programming: generics-sop. I’ve been playing with this library in the last couple of days, and I absolutely love the approach. In today’s short post, I want to demonstrate how easy it is to use this library. I don’t plan to go into a lot of detail, but I encourage interested readers to check out the associated paper - True Sums of Products - a paper with a lovely balance of theory and a plethora of examples.

postgresql-simple

When working with postgresql-simple, one often defines records and corresponding FromRow and ToRow instances. Let’s assume we’re modelling a library. No library is complete without books, so we might begin with a record such as:

data Book = Book
  { bookTitle :: Text
  , bookAuthor :: Text
  , bookISBN :: ISBN
  , bookPublishYear :: Int
  }

In order to store and retrieve these in our database, we need to write the following instances:

instance FromRow Book where
  toRow = Book <$> field <*> field <*> field <*> field

instance ToRow Book where
  toRow Book{..} =
    [ toField bookTitle
    , toField bookAuthor
    , toField bookISBN
    , toField bookPublishYear
    ]

As you can see - that’s a lot of boilerplate. In fact, it’s nearly twice as much code as the data type itself! The definitions of these instances are trivial, so it’s frustrating that I have to manually type the implementation bodies by hand. It’s here that we turn to generics-sop.

First, we’re going to need a bit of boiler-plate in order to manipulate Books generically:

data Book = ...
  deriving (GHC.Generics.Generic)

instance Generics.SOP.Generic Book

We derive generic representations of our Book using GHC.Generics, and in turn use this generic representation to derive the Generics.SOP.Generic instance. With this out of the way, we’re ready to work with Books in a generic manner.

generics-sop

The generics-sop library works by manipulating heterogeneous lists of data. If we look at our Book data type, we can see that the following two are morally describing the same data:

book = Book "Conceptual Mathematics" "Lawvere, Schanuel" "978-0-521-71916-2" 2009
book = [ "Conceptual Mathematics", "Lawvere, Schanuel", "978-0-521-71916-2", 2009 ]

Of course, we can’t actually write such a thing in Haskell - lists are required to have all their elements of the same type. However, using modern GHC extensions, we can get very close to modelling this:

data HList :: [*] -> * where
  Nil :: HList '[]
  (:*) :: x -> HList xs -> HList (x ': xs)

book :: HList '[Text, Text, ISBN, Int]
book = "Conceptual Mathematics"
    :* "Lawvere, Schanuel"
    :* "978-0-521-71916-2"
    :* 2009
    :* Nil

Once we begin working in this domain, a lot of the techniques we’re already familiar with continue fairly naturally. We can map over these lists, exploit their applicative functor-like structure, fold them, and so on.

generics-sop continues in the trend, using kind polymorphism and a few other techniques to maximise generality. We can see what exactly is going on with generics-sop if we ask GHCI for the :kind! of Book’s generic Code:

> :kind! Code Book
Code Book = SOP I '[ '[ Text, Text, ISBN, Int ] ]

The list of fields is contained within another list of all possible constructors - as Book only has one constructor, thus there is only one element in the outer list.

FromRow, Generically

How does this help us solve the problem of our FromRow and ToRow instances? First, let’s think about what’s happening when we write instances of FromRow. Our Book data type has four fields, so we need to use field four times. field has side effects in the RowParser functor, so we sequence all of these calls using applicative syntax, finally applying the results to the Book constructor.

Now that we’ve broken the problem down, we’ll start by solving our first problem - calling field the correct number of times. Calling field means we need to have an instance of FromField for each field in a constructor, so to enforce this, we can use All to require all fields have an instance of a type class. We also use a little trick with Proxy to specify which type class we need to use. We combine all of this with hcpure, which is a variant of pure that can be used to build a product:

fields :: (All FromField xs, SingI xs) => NP RowParser xs
fields = hcpure fromField field
  where fromField = Proxy :: Proxy FromField

So far, we’ve built a product of field calls, which you can think of as being a list of RowParsers - something akin to [RowParser ..]. However, we need a single row parser returning multiple values, which is more like RowParser [..]. In the Prelude we have a function to sequence a list of monadic actions:

sequence :: Monad m => [m a] -> m [a]

There is an equivalent in generics-sop for working with heterogeneous lists - hsequence. Thus if we hsequence our fields, we build a single RowParser that returns a product of values:

fields :: (All FromField xs, SingI xs) => RowParser (NP I xs)
fields = hsequence (hcpure fromField field)
  where fromField = Proxy :: Proxy FromField

(I is the “do nothing” identity functor).

Remarkably, these few lines of code are enough to construct data types. All we need to do is embed this product in a constructor of a sum, and then switch from the generic representation to a concrete data type. We’ll restrict ourselves to data types that have only one constructor, and this constraint is mentioned in the type below (Code a ~ '[ xs ] forces a to have only one constructor):

gfrowRow
  :: (All FromField xs, Code a ~ '[xs], SingI xs, Generic a)
  => RowParser a
gfrowRow = to . SOP . Z <$> hsequence (hcpure fromField field)
  where fromField = Proxy :: Proxy FromField

That’s all there is to it! No type class instances, no skipping over meta-data - we just build a list of field calls, sequence them, and turn the result into our data type.

ToRow, Generically

It’s not hard to apply the same ideas for ToRow. Recall the definition of ToRow:

class ToRow a where
  toRow :: a -> [Action]

toRow takes a value of type a and turns it into a list of actions. Usually, we have one action for each field - we just call toField on each field in the record.

To work with data generically, we first need move from the original data type to its generic representation, which we can do with from and a little bit of pattern matching:

gtoRow :: (Generic a, Code a ~ '[xs]) => a -> [Action]
gtoRow a =
  case from a of
    SOP (Z xs) -> _

Here we pattern match into the fields of the first constructor of the data type. xs is now a product of all fields, and we can begin turning into Actions. The most natural way to do this is simply to map toField over each field, collecting the resulting Actions into a list. That is, we’d like to do:

map toField xs

That’s not quite possible in generics-sop, but we can get very close. Using hcliftA, we can lift a method of a type class over a heterogeneous list:

gtoRow :: (Generic a, Code a ~ '[xs], All ToField xs, SingI xs) => a -> [Action]
gtoRow a =
  case from a of
    SOP (Z xs) -> _ (hcliftA toFieldP (K . toField . unI) xs)

  where toFieldP = Proxy :: Proxy ToField

We unwrap from the identity functor I, call toField on the value, and then pack this back up using the constant functor K. The details here are a little subtle, but essentially this moves us from a heterogeneous list to a homogeneous list, where each element of the list is an Action. Now that we have a homogeneous list, we can switch back to a more basic representation by collapsing the structure with hcollapse:

gtoRow :: (Generic a, Code a ~ '[xs], All ToField xs, SingI xs) => a -> [Action]
gtoRow a =
  case from a of
    SOP (Z xs) -> hcollapse (hcliftA toFieldP (K . toField . unI) xs)

  where toFieldP = Proxy :: Proxy ToField

Admittedly this definition is a little more complicated than one might hope, but it’s still extremely concise and declarative - there’s only a little bit of noise added. However, again we should note there was no need to write type class instances, perform explicit recursion or deal with meta-data - generics-sop stayed out of way and gave us just what we needed.

Conclusion

Now that we have gfromRow and gtoRow it’s easy to extend our application. Perhaps we now want to extend our database with Author objects. We’re now free to do so, with minimal boiler plate:

data Book = Book
  { bookId :: Int
  , bookTitle :: Text
  , bookAuthorId :: Int
  , bookISBN :: ISBN
  , bookPublishYear :: Int
  } deriving (GHC.Generics.Generic)

instance Generic.SOP.Generic Book
instance FromRow Book where fromRow = gfromRow
instance ToRow Book where toRow = gtoRow

data Author = Author
  { authorId :: Int
  , authorName :: Text
  , authorCountry :: Country
  } deriving (GHC.Generics.Generic)

instance Generic.SOP.Generic Author
instance FromRow Author where fromRow = gfromRow
instance ToRow Author where toRow = gtoRow

generics-sop is a powerful library for dealing with data generically. By using heterogeneous lists, the techniques we’ve learnt at the value level naturally extend and we can begin to think of working with generic data in a declarative manner. For me, this appeal to familiar techniques makes it easy to dive straight in to writing generic functions - I’ve already spent time learning to think in maps and folds, it’s nice to see the ideas transfer to yet another problem domain.

generics-sop goes a lot further than we’ve seen in this post. For more real-world examples, see the links at the top of the generics-sop Hackage page.


You can contact me via email at ollie@ocharles.org.uk or tweet to me @acid2. I share almost all of my work at GitHub. This post is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License.