Solution to servant API type indentation problem

Posted on 2017-03-15 by Oleg Grenrus servant

This post is a literate Haskell file: so there are few imports, and {-# LANGUAGE TypedKitchenSink #-}.

{-# LANGUAGE DataKinds #-}
{-# LANGUAGE KindSignatures #-}
{-# LANGUAGE OverloadedStrings #-}
{-# LANGUAGE RecordWildCards #-}
{-# LANGUAGE ScopedTypeVariables #-}
{-# LANGUAGE TemplateHaskell #-}
{-# LANGUAGE TypeFamilies #-}
{-# LANGUAGE TypeOperators #-}
{-# OPTIONS_GHC -Wall #-}
module ListApi where
import Data.Text (Text)
import Data.Type.Equality
import Generics.SOP
import Generics.SOP.TH
import Servant

#Problem

How you indent your servant api types?

type API1 = "foo" :> Get '[JSON] Int
       :<|> "bar" :> QueryParam "q" Text :> Get '[JSON] Text
       :<|> "quu" :> ReqBody '[JSON] Text :> Post '[JSON] Bool

type API2 =
    "foo" :> Get '[JSON] Int
    :<|> "bar" :> QueryParam "q" Text :> Get '[JSON] Text
    :<|> "quu" :> ReqBody '[JSON] Text :> Post '[JSON] Bool

type API3 =
         "foo" :> Get '[JSON] Int
    :<|> "bar" :> QueryParam "q" Text :> Get '[JSON] Text
    :<|> "quu" :> ReqBody '[JSON] Text :> Post '[JSON] Bool

There seems to be an unsolvable aesthetic vs. diff-friendly problem!

#Solution

Let's add a type-family!

type family ListApi (es :: [*]) :: * where
    ListApi '[e]      = e
    ListApi (e ': es) = e :<|> ListApi es

With the help of ListApi, we can reduce the indentation problem to how we indent (type-level) lists:

type API' =
   '[ "foo" :> Get '[JSON] Int
    , "bar" :> QueryParam "q" Text :> Get '[JSON] Text
    , "quu" :> ReqBody '[JSON] Text :> Post '[JSON] Bool
    ]

type API = ListApi API'

The API and previously defined types are the same:

proof :: API :~: API1
proof = Refl

#Follow-up

Yet to write the Server API, we still need to use :<|>. The problem is not solved properly yet. There is generic client in servant-client (PR#640), and we can use the same idea for server part too: we need one more type family and generics-sop:

type family ServerMap (xs :: [*]) :: [*] where
    ServerMap '[]       = '[]
    ServerMap (x ': xs) = Server x ': ServerMap xs

newtype Ser x = Ser (Server x)

listApiServer
    :: forall a x xs.
       (Generic a, SListI xs, Code a ~ '[ServerMap (x ': xs)])
    => Proxy (x ': xs) -> a -> Server (ListApi (x ': xs))
listApiServer _ =
    step1 . step0 (shape :: Shape (x ': xs)) . unZ . unSOP . from
  where
    step0 :: forall ys. Shape ys -> NP I (ServerMap ys) -> NP Ser ys
    step0 ShapeNil      Nil         = Nil
    step0 (ShapeCons s) (I x :* ys) = Ser x :* step0 s ys

    step1 :: forall y ys. NP Ser (y ': ys) -> Server (ListApi (y ': ys))
    step1 (Ser x :* Nil)         = x
    step1 (Ser x :* ys@(_ :* _)) = x :<|> step1 ys

    -- this doesn't work:
    -- inAsingleStep
    --     :: Shape ys
    --     -> NP I (ServerMap (y ': ys))
    --     -> Server (ListApi (y ': ys))
    -- inAsingleStep ShapeNil (I x :* Nil) = Nil

This function is written in the infamous banzai mode, to convince GHC the program is well-typed. The Proxy argument is not obviously required, but the following code would be ambiguous without it. The Shape parameter in step0 is needed because GHC cannot reason backwards ServerMap xs ~ '[]xs ~ '[] (ServerMap is injective only shallowly, Server isn't injective!). We need two steps, because going directly from NP I (ServerMap ys) to Server (ListApi ys) would require to deal with two type-families at once, and GHC had Also, step1 highlights the question: should we have a type for an identity element of :<|> operation?

#An example

With the help of ListApi and listApiServer, we could define the server implementation in a neat way, IMO:

api :: Proxy API
api = Proxy

api' :: Proxy API'
api' = Proxy

data ApiImpl m = ApiImpl
    { apiFoo :: m Int
    , apiBar :: Maybe Text -> m Text
    , apiQuu :: Text -> m Bool
    }

deriveGeneric ''ApiImpl

check :: Code (ApiImpl Handler) :~: '[ServerMap API']
check = Refl

app :: Application
app = serve api $ listApiServer api' impl
  where
    impl :: ApiImpl Handler
    impl = ApiImpl {..}

    apiFoo = return 42

    apiBar Nothing  = return "empty"
    apiBar (Just x) = return x

    apiQuu t = return (t == "works?")

If you experiment with the code and try to change the types in ApiImpl fields, then the check "test" will catch them. It seems, that errors generated from type-checking check are way more to the point than type-errors in "classic" serve api $ endpoint1 :<|> endpoint2 :<|> ... approach.

I tend to have small wrappers around actual implementation of endpoints, (i.e. I do write apiFoo which calls businessFoo, which knows little or nothing about the web) so the ApiImpl is the only additional boilerplate.

Would this be useful? Give it a spin! Please comment either directly to me (tweet) or to servant issue.


You can run this file with

stack --resolver=nightly-2017-03-01 ghci --ghci-options='-pgmL markdown-unlit'
*Main> :l servant-server-type.lhs

fetch the source from https://gist.github.com/phadej/06e0ce32e790fdab63b119fcbad357f5

P.S. Someone, please fix DataKinds syntax highlighting in whatever pandoc uses!

Site proudly generated by Hakyll