A Batch Querying Applicative Functor Without IORefs

Regular readers of my blog will remember that I recently blogged about a Querying applicative functor, which is able to “watch” for queries to a data store and batch multiple queries into one. It’s able to do all of this while allowing us to write code in a declarative manner, using traversals and other techniques we’re already familiar with. To refresh your memory, here’s an example what we can write with Querying:

withQuery getUsersById $ \usersById ->
withQuery getTagsByPostId $ \tagsByPostId ->
  for blogPosts $ \post ->
    BlogPost <$> pure post
             <*> usersById @? postAuthor post
             <*> tagsByPostId @? postId post

In this example, we have a list of blogPosts that we want to load more information for. We loop over each post, and load the author and tags each post. Despite using a loop, by using Querying we only perform two queries, as Querying batches the queries together.

Even though we reached a viable solution, it left me unsatisfied. The solution relied on mutability (twice!) - once to record which keys were looked up (here we used an IORef), and again to create a suspended computation containing the result (an initially empty MVar). There’s no reason we should have to require mutability, so I’ve been searching for a pure solution - and I’m happy to say I’ve found it.

One at a Time, Please

To start understanding the problem, let’s look at a pure Querying applicative that works with a single query. A lot of people on Reddit immediately suggested this as an alternative solution to my IORef solution, but didn’t realise it only works for a single query. Nonetheless, it’s a good starting point:

data Querying k v a = Querying [k] ([(k, v)] -> a)

This applicative functor can be thought of as the product of two smaller applicative functors - the Const [k] functor collects a list of keys, and the ([(k, v)] ->) functor allows us to create a suspended computation awaiting results. If we write out the applicative instance by hand, we can see what it means to combine two Queryings together:

instance Applicative (Querying k v) where
  pure x = Querying mempty (pure x)
  (Querying ks f) <*> (Querying ks' x) =
    Querying (ks <> ks') (\result -> f result <*> x result)

The pure computation asks no questions and thus doesn’t care what the results of the query are. We can combine two Queryings together by collecting the keys of both values, and creating a new computation suspend on the result that combine the underlying values.

The problem is that this applicative is now parameterized on a single key/value pair. If we have a Querying Int String and a Querying Int Bool, the two are different types and thus we can’t use <*> to combine them.

What we need to do is be able to store lists of key value pairs of potentially different types. If we could do this, then we can have a list of keys for every key value pair. Likewise, we want a similar construction for our suspended computation, one that would depend on the results of all queries. If we ask 3 queries, then we would like a function of three arguments.

KeyLists

First of all, let’s see how we can have a list of keys of different types. Since GHC learned how to promote data types to the type level, things have become a lot easier here. First, we’ll introduce a useful type of key value pairs - this will help us stay organised and principled at the type level:

data KV k v = KV k v

We won’t actually be constructing any terms of type KV, but we need a constructor to promote to be a type. Next, we introduce a GADT KeyList which is a list of lists of keys:

data KeyList :: [KV * *] -> * where
  Nil :: KeyList '[]
  Cons :: [k] -> KeyList kvs -> KeyList ('KV k v ': kvs)

The empty key list contains no information, and by constructing a term with Nil, we learn that there are no keys at all - as we can see by the empty list in the type of Nil. However, we also have a Cons constructor which takes a list of keys (of type k), and appends this list onto another list of potentially different types. So here we can see that we combine terms with cons, and we mirror this structure at the type level - we are consining a new KV type in the final type of Cons.

To emphasise the list-like nature of this data type, we can make it an instance of Monoid, which will be useful later. However, there are actually infinitely many monoid instances - one for every possible list of KVs. Fortunately, Haskell’s type classes are able to deal with this, by generating Monoid instances out of smaller ones:

instance Monoid (KeyList '[]) where
  mempty = Nil
  mappend _ _ = Nil

instance (Monoid (KeyList kvs)) => Monoid (KeyList ('KV k v ': kvs)) where
  mempty = Cons mempty mempty
  mappend (Cons l ls) (Cons r rs) = Cons (l ++ r) (mappend ls rs)

These monoid instances are somewhat reminiscent of an inductive proof - our base case is the empty list, and a monoid instance for a non-empty list relies on the monoid instance for a one-element-smaller list, which is a lot like making use of an inductive hypothesis.

An Answer For Every Question

Now that we know that we can work with lists of keys of different types, we should next spend some time to work out how we’re going to create the suspended computation for multiple questions. Let’s look at a few examples to get a feel for things. For a single query, we want:

[(k, v)] -> a

For two queries, we want:

[(k1, v1)] -> [(k, v)] -> a

Hmm, what about zero queries? Well, logically this would just be a term of type a - no function at all. Well hey, this again looks a lot like a list! We can use a very similar technique to building KeyList - but instead of storing lists, we would store functions. The constant value moves to be part of our base case:

data ResultF :: [KV * *] -> * -> * where
  ResultConst :: a -> ResultF '[] a
  ResultFunc :: ([(k, v)] -> ResultF kvs a) -> ResultF ('KV k v ': kvs) a

Because the base case carries a term, we need a way to mention the type of this term, so ResultF kvs actually constitutes a Functor:

instance Functor (ResultF kvs) where
  fmap f (ResultConst x) = ResultConst (f x)
  fmap f (ResultFunc g) = ResultFunc (fmap f . g)

This is a fairly natural definition - we just pattern match and either apply our function to the constant value, or we fmap inside the function, which builds up a new function.

In fact, ResultF actually creates a full applicative functor - a hint that we’re on the right track to building a larger applicative functor out of it. Applicative functors have the pure method as part of their type class, which is a producer of data. As such, we’ll have to make sure we have instances for any possible type of ResultF, as we did with our KeyList Monoid:

instance Applicative (ResultF '[]) where
  pure = ResultConst
  (ResultConst f) <*> (ResultConst x) = ResultConst (f x)

instance Applicative (ResultF kvs) => Applicative (ResultF ('KV k v ': kvs)) where
  pure x = ResultFunc (const (pure x))
  (ResultFunc f) <*> (ResultFunc x) = ResultFunc (\args -> f args <*> x args)

These aren’t particularly pretty, but fortunately they aren’t particularly surprising either.

Believe it or not, we’re now ready to build our Querying functor!

data Querying :: (* -> *) -> [KV * *] -> * -> * where
  Querying :: KeyList kvs -> Compose m (ResultF kvs) a -> Querying m kvs a

instance Functor m => Functor (Querying m kvs) where
  fmap f (Querying kvs x) = Querying kvs (fmap f x)

instance (Applicative m, Applicative (ResultF kvs), Monoid (KeyList kvs)) =>
           Applicative (Querying m kvs) where
  pure a = Querying mempty (pure a)
  (Querying kvs f) <*> (Querying kvs' x) = Querying (kvs <> kvs') (f <*> x)

Querying itself is just a combination of a key list and a monadic action producing a ResultF that eventually produces an a (I use Compose to help me write the Applicative instances). The Functor instance leaves the key list un-changed, and just uses the Functor instance for ResultF to do the heavy lifting.

The applicative instance may look a little bit more intricate, but look at the definitions… They are exactly the same as what we had before! This shouldn’t be surprising, but somehow… I can’t help but be surprised :)

Query Pointers

Now that we have the ability to build up KeyList and ResultF terms we’re making good progress. I have a loose plan - I want withQuery to introduce another level of scope in our Querying applicative. This corresponds to Cons for KeyList and adding a new ResultFunc layer for ResultF. Introducing a new level of scope will provide the user with a pointer to this scope, which gives them a way to demand results from the query.

But how do we represent these pointers? For usual lists, we can use natural numbers, but if we used numbers here we’d throw an awful lot of useful information away. A better description would be to describe the operations that someone has to do to the current scope to get to the data they seek. There are two possibilities: either the data you seek is at the outermost level of scope, or you have a pointer that is valid under that scope. You can think of this as “stripping away” one level of scoping and then carrying on a lookup. It may be easier to understand this by seeing the data type and some examples:

data Where :: [KV * *] -> (KV * *) -> * where
  Here :: Where (kv ': tail) kv
  There :: Where kvs kv -> Where ('KV k v ': kvs) kv

Where carries a list of scopes that it’s valid in as its first parameter, and its second parameter indicates exactly which key-value pair we’re pointing to. Either we’re pointing to the outermost scope with Here, or we want to strip off one layer of scope and then carry on.

For example, we could construct the following queries:

Here :: '['KV Int Int] ('KV Int Int)
There Here :: '['KV Int String, 'KV Int Int] ('KV Int Int)
There Here :: '['KV String Char, 'KV Int String, 'KV Int Int] ('KV Int String)

Introducing Queries

The Where data type will provide us with a way to introduce a pointer to a query which is useful for withQuery. If I tease you with the type of withQuery, it will be easier to see what the plan is:

withQuery
  :: ([k] -> m [(k, v)])
  -> (Where ('KV k v ': kvs) ('KV k v) -> Querying m ('KV k v ': kvs) a)
  -> Querying m kvs a

withQuery will take a query in some monad m, and a function to operate in a larger scope. We can see that we extend an existing Querying of kvs with a new level of scope for KV k v. The pointer that we pass in only has to reference the outermost scope, and we know that we can do that with Here. When we call the continuation with Here, we get back a Querying that contains some keys and a ResultF. withQuery then performs the query, and removes this outer scope to get us back to where we were:

withQuery query k =
  case k Here of
    Querying (keys `Cons` kvs) (Compose m) ->
      Querying kvs $ Compose $ do
        ResultFunc f <- m
        results <- query keys
        return (f results)

Once we have a query, how do we actually ask questions? A first approach might look like this:

ask :: (Eq k, Monad m) => Where kvs ('KV k v) -> k -> Querying m kvs (Maybe v)
ask q k = Querying (mkKeyList q k) (Compose $ return (mkResultF q k))

mkKeyList :: Where kvs ('KV k v) -> k -> KeyList kvs
mkKeyList Here k = Cons [k] (mempty)
mkKeyList (There path) k = Cons [] (mkKeyList path k)

mkResultF :: (Eq k) => Where kvs ('KV k v) -> k -> ResultF kvs (Maybe v)
mkResultF Here k = ResultFunc (identityResultF . lookup k)
mkResultF (There path) k = ResultFunc (const (mkResultF path k))

This is a reasonable amount of code, so lets walk through it. When we ask a query for the value under a specific key, we need to do two things. First we have to record that we are looking up a specific key, and secondly we have to create a suspended computation to lookup the result. Both of these are built by looking at the Where value for the query, which tells us how the index into the KeyList for keys for this query, and the corresponding position in the ResultF functor to extract the results.

mkKeyList and mkResultF simply iterate on a Where value until they find Here. Once they do, they simply return a KeyList or ResultF from that point all the way to the bottom. For KeyList that’s easy - we use the Monoid instance to create empty cells for all subsequent nodes. For ResultF it’s a little more complex - but the idea is we simply ignore any subsequent parameters and just return our result. Check out the full code listing for a definition of identityResultF.

If we write a way to actually run a Querying, then what we have is already useful:

runQuerying :: Monad m => Querying m '[] a -> m a
runQuerying (Querying Nil (Compose m)) = liftM (\(ResultConst a) -> a) m

example :: IO [(Maybe String)]
example =
  runQuerying $
    withQuery getUserNameById $ \userNameById ->
      for [1..10] $ \userId ->
        ask userNameById userId

  where
  getUserNameById ids =
    print ids
    return [(1, "Alice"), (2, "Bob")]
> example >>= print

Cool!

However, things fall apart pretty quickly if we try two queries at once:

example :: IO [(Maybe String, Maybe Int)]
example =
  runQuerying $
    withQuery getUserNameById $ \userNameById ->
    withQuery getUserAgeById $ \userAgeById ->
      for [1..10] $ \userId ->
        (,) <$> ask userNameById userId
            <*> ask userAgeById userId

So, what’s gone wrong? The problem is that when we call withQuery we are given a path into a specific list of key-values. However, withQuery modifies this very list. Thus if we open two queries, then the first query is no longer valid - because the second query is now operating in a larger environment. What we need to do is extend our old references to work with larger scopes. How would we do that?

When we extend the scope of a query, we grow the list of keys and the list of arguments in ResultF by one. Thus to make an existing query valid in this environment, we need to prepend There onto the query. Remember that There simply skips one layer - in this case stripping off a layer of scope and getting back to where we were.

It’s completely mechanical how many times we need to add There, and we want to exploit this and have GHC figure it out for us. This means we need to infer the amount of Theres to add, and this is a perfect job for type classes. I’m going to add a new type class, Somewhere, which will take a Where term in one environment, and give you an equivalent path in a larger environment.

class Somewhere (kvs :: [KV * *]) (kvs' :: [KV * *]) (kv :: (KV * *)) where
  somewhere :: Where kvs kv -> Where kvs' kv

All that’s left is to write the instances. The trivial case is that you are you trying to use a query in an environment where the query is already at the outer-most scope. In this case, the path is simply Here:

instance Somewhere ('KV k v ': kvs') ('KV k v ': kvs') ('KV k v) where
  somewhere _ = Here

Notice how we are referring to the same k and v at every place in the context for this type class instance.

The other scenario is that we have an environment that’s too big. In this case we use There, and try again with the smaller environment:

instance Somewhere kvs kvs' kv => Somewhere kvs ('KV k v ': kvs') kv where
  somewhere = There . somewhere

These instances overlap, so we’ll need -XOverlappingInstances, but it seems to be safe here.

Armed with this type class, we just need to modify ask to call somewhere, which now makes it more polymorphic in the environment type:

ask :: (Eq k, Monad m, Somewhere kvs kvs' ('KV k v)) => Where kvs ('KV k v) -> k -> Querying m kvs' (Maybe v)
ask q k = Querying (mkKeyList (somewhere q) k) (Compose $ return (mkResultF (somewhere q) k))

Again, notice that the query kvs is not necessarily the same as Querying’s kvs'. We require that the two environments are compatible though by Somewhere to the context of this function.

With this last change, we’re done - our example now compiles:

Concluding Thoughts

When I first attempted to work on the batching Querying applicative functor, I felt like it could obviously be done with a pure solution. However, after a day of dead ends, I moved over to the IORef solution. However, as this post shows - the solution is very much possible and doesn’t require a particularly different formulation. What the solution does require is the ability to really understand how various different techniques (such as GADTs and type class) can be used, and how to retain information within the type system. The moment you lose information, problems of this nature become very complicated.

While I’ve provided a (hopefully!) coherent solution above, by no means was it as simple as I may have made out! Infact, this problem is a perfect case study of my favourite type of problems - the class of problems that are just outside my current capabilities. I highly encourage everyone to hold on to these problems - they don’t come up often, but the rewards if you follow it all the way to the solution are substantial. Don’t give up, but don’t be afraid to ask questions either.


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.