A server-client architecture

In this installment of the series, we’ll see:

  • how to implement a client-server architecture, with a common package to share code and abstractions between the two parts.
  • how to use the package servant-reflex to seamlessy embed server requests in the frp network.
  • how to use a library to talk about data validation, of the kind done in html forms.

The code for today’s repo is in: TODO

Let’s begin with the simplest matter: how to share data definitions and abstractions between the backend and the frontend. It seems a very widespread practice to create three packages: one, let’s say common, will contain the shared abstractions, and will be included by the other two, client (with the code for the webapp, to be compiled with ghcjs), and server (with the code for the server, to be compiled with ghc). That’s all.

Let’s also briefly describe here what this application does and the structure of the server: TODO


The requisites for validation

When designing a web app there are two kinds of validations that can be run: the first is the one done on the client, to provide validation against crude error (think of inputing a well-formed email address); the other one, usually done on the server, is about validating the data against our knowledge (think of checking if an email address is in the user database).

Sometimes, for security reasons, the server might want to do again the validations which happened in the client, and so we need way of easily composing validations, sharing the common infrastructure, so that code duplication is reduced.

Another problem that we encouter is that the format in which we report back the error to the client must be convenient enough to report errors near the UI element which caused them; for example, when validating a user with a combination of mail and password, an error message for a wrong password should be displayed near the password itself.

This brings us to discussing common solution for validation: there is the Data.Validation approach, in the validation package, which is essentially Either with another applicative instance. Unfortunately this approach fails us because we have no obvious way of reporting back errors to their use site.

On the other hand we have the digestive-functors approach, which unfortunately is geared towards a server-centric approach, and makes validations on the client difficult to write (TODO: Check the correctness of this information with Jasper).

A possible solution

So let’s think about another solution: let’s say I’m implementing a Mail/Password validation, so the type of my user could be

data User = User Mail Text

Now, if we expand slightly our definition to

data UserShape f = UserShape (f Mail) (f Text)

we gain the possibility of talking about a structure whose fields talk about operations or data parametrized by Mail and Text.

For example, some functor that we might want to use are Identity (and in fact User is obiously isomorphic to UserShape Identity), Maybe or Either Text to model the presence of errors, or for example

newtype Validation a = Validation { unValidationF :: Compose ((->) a) Maybe a }

so that:

UserShape Validation ~ UserShape (Mail -> Maybe Mail) (Text -> Maybe Text)

Now that we can talk about this “user shaped” objects, we might want to combine them, for example with something like:

validateUser :: User -> UserShape Validation -> UserShape Maybe

the shaped library has a generic mechanism of doing this kind of manipulations (check out the validateRecord function). The library uses internally generics-sop to construct and match the generic representations, and some Template Haskell to shield the user from the boilerplate instance declarations.

Now, we can send to the server a tentative User to check, and get back a UserShape Maybe that we can easily map back to our input text-boxes.

You can check how that’s done in the client for today’s installment (TODO link the correct lines).

How to query the API endpoint

The common code in this simple case contains only the definition of the user type and the type for our servant API

The server code is a simple server that serves a mock authentication. I’m not entering in an in depth discussion on the servant approach here (if you’re interested check the wonderful servant documentation, but the gist is that you can create from a description of the api, in this project:

type MockApi = "auth" :> ReqBody '[JSON] User :> Post '[JSON] Text
             :<|> Raw

A server satisfying that api, here:

server :: Server MockApi
server = authenticate :<|> serveAssets :<|> serveJS

The package servant-reflex transforms a Servant API in Reflex functions for querying it, in the same way servant-server transforms it in a server. The invocation is very easy:

let url = BaseFullUrl Http "localhost" 8081 ""
(invokeAPI :<|> _ :<|> _) = client (Proxy @MockApi) (Proxy @m) (constDyn url)
client :: HasClient t m layout => Proxy layout -> Proxy m -> Dynamic t BaseUrl -> Client t m layout

As you can see, client is the most important function: it takes proxies for the API and the monad in which the computation is executed (as it’s customary to run a reflex computation in a (constrained) universally quantified monad, like our own body :: MonadWidget t m => m () (the syntax with @ is due to the ghc 8’s TypeApplications extension, without it you should have written Proxy :: Proxy MockApi etc.)

That gives us a mean to call the relevant API endpoint (TODO: detail the type of the transformed function, detailing how the API call is translated in events. Also talk about Xhr).

For example in our code we use this feature to like this: