Posts Tagless Final Encoding in Haskell

Tagless Final Encoding in Haskell

Source Code

You can find source code of the example described in this post here


In this post i am going to explore a simple technique for organizing our programs which is called Tagless Final Encoding to write testable programs in Haskell. I am also use TypeApplication LANGUAGE directive to write more readable and flexible test.

Why Tagless Final?

Nowadays in Haskell Community there is an open discussion about using Free Monads, mtl or Tagless Final Encoding to write internal DSL (Domain Specific Language) for representing our programs in a descriptive and Functional way.

In my personal opinion i think all of these tools, theories and techniques are suitable to do it but depends on the context of the person, team or solution you are writing to decide which is more useful.

For example:

  • Free Monad: I think it is great to have tools that are based on Category Theory concepts such as Free Applicative, Free Monad and so on. There is a great paper about this Notions of Computation as Monoid. In that sense Free Monad not only help us to describe our programs but also to have certain Math Properties in our toolbox to manipulate them. Although i have never benchmarked any of the Free Monad implementations out there, i know there are complains about their performance in the community. Beyond this i think for beginners it is a little difficult to implement.

  • Monad Transformers (mtl): Also great tool, based on simple Monad concept which is easier to understand for beginners and without performance penalties if you are using carefully. It is also the most used tool for dealing with different Monads in a single program from the beginning of Haskell. The only disadvantage i could pointed out is that is less readable and understandable in the code when you are stacking more than 3 Monads. Also a drawback for beginners.

  • Tagless Final: You only need to define and implement Typeclasses. In the original post i have wrongly mentioned “it is a technique not based on any paper or Math Theory”, but thank you to p-alik who pointed me out in reddit channel that there is paper for this here Typed final (tagless-final) style. The main advantage for me is it is beginner friendly, readable, easy to understand, test and extend.

Having said that, I would like to talk about Tagless Final as an approach for Haskell beginners in order to help them to organize and describe programs; make them extensible and testable.

What is Tagless Final Encoding?

Tagless Final Encoding is a technique for embedding a DSL (Domain Specific Language) in a Functional Programming Language. We need to define a Language for using it and an Interpreter to indicate how it should behave on each defined term. For this purpose we are going to use Typeclasses.

To sum up in Tagless Final Encoding style there are:

  • Typeclasses: Set of operations over a Type.

  • Interpreter: Instances of those Typeclasses for each specific Type

Tagless Final Encoding in practice

We are going to build a basic program which request some user data. The program is going to do the following:

  • First try to recover the data from a cache
  • If data is found it is returned
  • If there is no data in the cache, search user data in source repository and update cache
type UserName = String

data DataResult = DataResult String
  deriving (Eq, Show)

requestData :: Monad m => UserName -> m [DataResult]
requestData userName = do
 cache  <- getFromCache userName
 result <- case cache of
   Just dataResult -> return dataResult
   Nothing         -> getFromSource userName
 storeCache result
 return result

Here it is our basic program which implements what we described above, but obviously this code doesn’t work because we need to define functions such as getFromCache, getFromSource and storeCache.

For defining that we are going to use Typeclasses as we mentioned, in order to represent our program capabilities.

class Monad m => Cache m where
  getFromCache :: String -> m (Maybe [DataResult])
  storeCache :: [DataResult] -> m ()

class Monad m => DataSource m where
  getFromSource :: String -> m [DataResult]

Why are we defining Cache and DataSource Typeclasses as Monad also? Basically because we want to combine and chain our DSL terms in a single program.

But we still need to change our program definition since we are constraining only on Monad and we want to use Cache and DataSource terms from the implicit context.

requestData :: (Cache m, DataSource m) => UserName -> m [DataResult]
requestData userName = do
 cache  <- getFromCache userName
 result <- case cache of
   Just dataResult -> return dataResult
   Nothing         -> getFromSource userName
 storeCache result
 return result

Notice that we don’t need anymore Monad Constraint in our signature because both Cache and DataSource are Monads also.

The only thing left is to write our Instances to provide some implementation. We are going to provide a fake implementation for IO.

instance Cache IO where
  getFromCache _ = return Nothing
  storeCache _ = return ()

instance DataSource IO where
  getFromSource user = return $ [DataResult $ "source: " <> user]

If we run our program from ghci we are going to see it is working:

λx.x> import Data

λx.x> requestData "john"
[DataResult "source: john"]

Provide and Test with different implementations using Type Application

One of the things I have announced on the beginning of this post is i am going to show how easy it is to test our programs using this technique combined with TypeApplication LANGUAGE extension. This combination enable us not only to test, but also to provide and interchange different instances of our Typeclasses in a straightforward way.


In order to provide different instances of Cache and DataSource, and play around with different cases, for example when data is in cache or not, i am going to wrappe IO type in different newtype representations.

{-# LANGUAGE GeneralisedNewtypeDeriving #-}

newtype NotInCache a = NotInCache { unNoCache :: IO a }
  deriving (Monad, Applicative, Functor)

instance Cache NotInCache where
  getFromCache _ = NotInCache $ return Nothing
  storeCache _ = NotInCache $ return ()

instance DataSource NotInCache where
  getFromSource user = return $ [DataResult $ "source: " <> user]

For the first instance we need to do enable GeneralisedNewtypeDeriving extension to allow us deriving Functor, Monad and Applicative because our Typeclasses Cache and DataSource are also Monad and we need to provide implementations of those Typeclasses for our custom type NotInCache

Now if we are trying to run this in ghci we are getting the following error:

λx.x> requestData "john"

<interactive>:5:1: error:
    • Ambiguous type variable ‘m0’ arising from a use of ‘print’
      prevents the constraint ‘(Show
                                  (m0 [DataResult]))’ from being solved.
      Probable fix: use a type annotation to specify what ‘m0’ should be.
      These potential instances exist:
        instance (Show a, Show b) => Show (Either a b)
          -- Defined in ‘Data.Either’
        instance Show a => Show (Maybe a) -- Defined in ‘GHC.Show’
        instance (Show a, Show b) => Show (a, b) -- Defined in ‘GHC.Show’ 14 others 89 instances involving out-of-scope types
        (use -fprint-potential-instances to see them all)
    • In a stmt of an interactive GHCi command: print it

Basically the compiler is saying us that it cannot find an unambiguous instance to use for our program. But also as the compiler is pointed out we can use TypeApplication extension to tell the compiler what instance should use and provide an explicit evidence of that.

λx.x> :set -XTypeApplications

λx.x> :t requestData "john"
requestData "john"
  :: (Data.Cache m, DataSource m, Data.Logging m) => m [DataResult]
λx.x> :t requestData @NotInCache "john"
requestData @NotInCache "john" :: NotInCache [DataResult]

Here we’ve enabled extension and after that we are running our program with NotInCache type. Notice that now we need to call unNoCache to unwrap our underlying IO and effectively running in our ghci IO loop.

x.x> unNoCache $ requestData "john"
[DataResult "source: john"]

We can also do it from our .hs file.

main :: IO ()
main = (unNoCache $ requestData "john") >>= (putStrLn . show)

Now we are ready for different instances!!!

newtype InCache a = InCache { unInCache :: IO a }
  deriving (Monad, Applicative, Functor)

instance Cache InCache where
  getFromCache user = InCache $ return $ Just [DataResult $ "cache: " <> user]
  storeCache _ = InCache $ return ()

instance DataSource InCache where
  getFromSource _ = undefined

main :: IO ()
main = do
  (unNoCache $ requestData "john") >>= (putStrLn . show)
  (unInCache $ requestData "john") >>= (putStrLn . show)

The outputs now look like this:

λx.x> Data.main
[DataResult "source: john"]
[DataResult "cache: john"]


One of the important aspects of Tagless Final Encoding is its extensibility property. It is extensible in 2 dimensions:

  • Vertical Extensibility: It is what we have just done adding different implementations for the same Typeclasses without altering our program.

  • Horizontal Extensibility: It is adding new capabilities to the program in order to extend some functionality inside it.

Horizontal Extensibility

Our program capabilities beyond Monad, Functor and Applicative Typeclasses are Cache and DataSource. If we are saying that it is Horizontal Extensible we can add more capabilities apart from those mentioned. For example what about Logging?.

Let do it with our example:

class Monad m => Logging m where
  logMsg :: String -> m ()

requestData :: (Cache m, DataSource m, Logging m) => UserName -> m [DataResult]
requestData userName = do
 cache  <- getFromCache userName
 result <- case cache of
   Just dataResult -> return dataResult
   Nothing         -> getFromSource userName
 storeCache result
 logMsg $ "Result data for user: " <> userName <> " - data: " <> show result
 return result

And now providing instances for Logging

instance Logging NotInCache where
  logMsg = NotInCache . putStrLn

instance Logging InCache where
  logMsg = InCache . putStrLn

If we run the program we obtain the following:

λx.x> Data.main
Result data for user: john - data: [DataResult "source: john"]
[DataResult "source: john"]
Result data for user: john - data: [DataResult "cache: john"]
[DataResult "cache: john"]


As we can see, Tagless Final Encoding is a pretty good technique to build testable and extensible programs.

We have also demonstrated how easy is to interchange and provide different instances using TypeApplication extension.

This post is licensed under CC BY 4.0 by the author.

Tagless Final in Scala: A Practical example


Comments powered by Disqus.