Your form has successfully submitted

One of our teammates will get back to you soon.

Using Cloud Haskell to write a type-safe distributed chat

In this tutorial we will implement a distributed chat with Erlang’s style and Haskell’s type safety using Cloud Haskell.

What is Cloud Haskell?

Cloud Haskell is a set of libraries that combines the power of Haskell’s type system with Erlang’s style of concurrency and distributed programming. This brings a lot of power in order to implement networking applications that are fault tolerant, can work on several network transports, and have an expressive model of communication. By expressive we mean that processes communicate explicitly by sending messages to each other rather than by sharing resources in memory. This also implies that processes -running in the same local node or in a remote node- pattern match over the specific messages they can handle, which fits very well with Haskell’s capabilities of modeling messages with algebraic data types.

Why Cloud Haskell?

Programming concurrent and distributed applications is way too hard. On top of the several challenges that you face when designing software, you will also struggle with race conditions, deadlocks, bad communication protocols, network problems that are hard to detect and to recover from, and code that is hard to debug and to maintain. That’s why Erlang was invented: Erlang is a language for manipulating distributed systems that focuses on recovery from failure.

Additionally, Erlang brought the possibility to write distributed programs in a functional style. All this is pretty interesting. However, it still lacks type-level guarantees since Erlang is a dynamically typed language, and we cannot model our concurrent programs as type-safe and predictable communication protocols. Cloud Haskell fills this gap by providing Erlang’s powerful distributed model shielded by Haskell’s powerful type system, so you can write your distributed programs with the robustness of Haskell and the error recovery from Erlang.


As an overview, let’s see how Cloud Haskell makes use of Erlang’s model by analyzing a very simple example. First, Cloud Haskell’s most fundamental entity is a process. Processes are isolated and lightweight threads of execution which run in a node, and the only way they can interact is by passing messages between each other. This is why processes are highly isolated since they do not share resources, which is the main cause of deadlocks and race conditions in distributed and concurrent systems. Keeping this in mind, sending a message to a process is as easy as creating a node for the process to reside in and sending a message to it with its unique process-id:

{-# LANGUAGE TupleSections     #-}

import Network.Transport.TCP (createTransport, defaultTCPParameters)
import Control.Distributed.Process
import Control.Distributed.Process.Node

main :: IO ()
main = do
  Right transport <- createTransport "" "4001" ("",) defaultTCPParameters
  node <- newLocalNode transport initRemoteTable
  _ <- runProcess node $ do
    -- get the id of this process
    self <- getSelfPid
    send self "Talking to myself"
    message <- expect :: Process String
    liftIO $ putStrLn message
  return ()

Even though the example above is not very interesting, it exposes some cool concepts about this platform: on the one hand, the process logic is decoupled from the network layer so that you can inject any transport backend you like. In our example, we are using createTransport to create a TCP/IP transport layer, and then we pass it to newLocalNode to create the node where our process will reside. Nevertheless, you could create a node with a different protocol like SSH, UDP, Unix Sockets, etc. On the other hand, as long as you have the unique identifier for a process, you can send a message to it. Thus, we can use getSelfPid to get the id of our process and then make it send the message "Talking to myself" to itself. After that, the process expects a message of type String which is queued in its mailbox. If no message of the expected type is in the mailbox, then the process will block until one arrives. In this case, the process is immediately receiving the "Talking to myself" string.

In this tutorial we will approach more interesting and advanced concepts of Cloud Haskell by implementing a simple chat server and client which supports the following features:

  • Launching a chat server room in a specific endpoint address that can be found by chat clients.
  • Launching chat clients that can search a chat server room in a specific endpoint address and connect to it.
  • Command line interface for writing messages in the chat room.
  • Broadcasting messages to all clients who are connected to a chat server room.
  • Handling the disconnection of clients from a chat server room.

First steps: The chat types

As mentioned above, one of the goals of Cloud Haskell is to set up an expressive messaging model between processes. By expressive we mean that we can specify which messages we want our process to handle, similar to specifying a communication protocol. This is achieved by pattern matching over the messages a process must handle and specifying some policy for unhandled messages which do not match any of the handlers. Thus, the first step in our implementation consists of defining the data types that will be signaled between our chat server and clients.

Initially, our chat server will have to handle two types of messages coming from clients, namely, 1) a message that lets a new client join the chat and 2) messages to be broadcasted to all the clients on the chat.

The first of these messages is straightforward:

{-# LANGUAGE DeriveDataTypeable     #-}
{-# LANGUAGE DeriveGeneric          #-}

module Types where

import GHC.Generics
import Data.Binary
import Data.Typeable
import Data.Map (Map)
import Control.Distributed.Process (SendPort)

type NickName = String

newtype JoinChatMessage = JoinChatMessage {
    clientName :: NickName
  } deriving (Generic, Typeable, Show)

instance Binary JoinChatMessage

Whenever a client wants to join a chat server, it provides a unique nickname that identifies him/her. Note that we are using Generics to derive instances for Binary and Typeable. This is necessary so that our type can automatically be an instance of Serializable which is about “objects that can be sent across the network” or, in other words, objects that can be encoded to raw bytestrings and decoded again to their original form.

Our second type represents any message which is broadcasted to the clients connected to a chat:

data Sender = Server | Client NickName
  deriving (Generic, Typeable, Eq, Show)

instance Binary Sender

data ChatMessage = ChatMessage {
    from :: Sender
  , message :: String
  } deriving (Generic, Typeable, Show)

instance Binary ChatMessage

Note that we capture the fact that the sender of that message can be either a client or the server itself -For example, when a client connects to the chat, the server broadcasts a message to the other clients announcing that a new member has joined.

Finally, with Cloud Haskell we can define processes which can update their state after handling a message. Thus, we can define the type of the state that the chat server process will update after a client joins:


import Control.Distributed.Process (SendPort)


type ClientPortMap = Map NickName (SendPort ChatMessage)

The state of our chat server process consists of a map from a client’s nickname (or identifier) to a send port. We’ll give more details about the SendPort data type when we talk about channels. Meanwhile, we can think about this type as an inventory of the clients that join the chat and a port through which we can send messages to them.

The server

As we mentioned above, the most fundamental entity in cloud Haskell is a process and that is why we can naturally define our chat server as a process. Nevertheless, in order to define server processes we can take advantage of the ProcessDefinition data type defined in the distributed-process-client-server package. A ProcessDefinition has several components which determine how different kinds of messages must be handled, but we will be focusing only on the two used in our chat server’s definition:

launchChatServer :: Process ProcessId
launchChatServer =
  let server = defaultProcess {
          apiHandlers =  [ handleRpcChan joinChatHandler
                         , handleCast messageHandler
        , infoHandlers = [ handleInfo disconnectHandler ]
        , unhandledMessagePolicy = Log
  in spawnLocal $ serve () (const (return $ InitOk M.empty Infinity)) server

Thus, a server process is basically a definition which specifies different kinds of handlers that match different types of messages. That’s why we have several kinds of handler lists in the ProcessDefinition, each one containing dispatchers that try to match a message queued in the process mailbox.

You can notice that our server process has two kinds of handlers, namely:

  • apiHandlers which are in charge of handling the core messages of the application, that is, messages from clients who want to join the chat, and messages which have to be broadcast to all the clients connected to the server.
  • infoHandlers which are useful for handling messages that clients are not explicitly sending to the server (e.g. when a client disconnects) and that have extra information about the SendPort which must be deregistered when a client disconnects.

Finally, we are specifying an unhandledMessagePolicy which makes the server log any of the messages which match none of the handlers defined above.

In order to have a better understanding of how our chat server will handle messages coming from the clients, let’s analyze the implementation of the handlers referenced in our process definition.

The server’s api handlers

Our chat server has a state represented by the ClientPortMap type which may be updated by a handler whenever this matches a specific message. Thus, besides matching specific messages, handlers also get access to the current state of the server which they can update according to the flow of the application. In our case, one of the handlers which must update the state of the application is the one in charge of registering the clients who connect to the chat server:

joinChatHandler :: ChannelHandler ClientPortMap JoinChatMessage ChatMessage
joinChatHandler sendPort = handler
    handler :: ActionHandler ClientPortMap JoinChatMessage
    handler clients JoinChatMessage{..} =
      if clientName `M.member` clients
      then replyChan sendPort (ChatMessage Server "Nickname already in use ... ") >> continue clients
        else do
          void $ monitorPort sendPort
          let clients' = M.insert clientName sendPort clients
              msg = clientName ++ " has joined the chat ..."
          logStr msg
          broadcastMessage clients $ ChatMessage Server msg
          continue clients'

ChannelHandler is a type synonym with the following definition:

type ChannelHandler state msg1 msg2 = SendPort msg2 -> (state -> msg1 -> Action state)

Which instantiated to the specific type parameters of the joinChartHandler definition, it would be:

  • state: The state of the server which is of type ClientPortMap
  • msg2: The type of message which can be sent through the SendPort, namely, a ChatMessage.
  • msg1: The type of message our handler is expecting from a client and that will be matched in the process' mailbox, that is, JoinChatMessage.

This definition expresses that it handles a channel by having as argument the SendPort of the chat client that is communicating to the server. This handler only matches messages of type JoinChatMessage and it replies to the clients with a message of type ChatMessage. A SendPort is one end of a tuple of communication ports whose other end is a ReceivePort. Together they compose an abstraction named channel which is useful for communicating two processes in a type-safe fashion. For example, in our handler, our chat server can only send messages of type ChatMessage to our clients through a port of type SendPort ChatMessage while the clients can only accept messages from this handler through a port of type ReceivePort ChatMessage.

With the concept of channel in mind, it is now clear that our handler basically does two things: if the client that is attempting to join the chat server wants to use a nickname that has already been taken by another client, the server will reply through its specific SendPort with a message notifying that the nickname is already in use ("Nickname already in use ..."). On the other hand, if the nickname is available, the server will broadcast a message to the currently connected clients notifying that a new user has joined the chat. The definition of the broadcast function is the following:

broadcastMessage :: ClientPortMap -> ChatMessage -> Process ()
broadcastMessage clientPorts msg =
  forM_ clientPorts (flip replyChan msg)

Which simply iterates over the ReceivePorts of the clients stored in the current server's state sending a ChatMessage with the replyChan function.

Finally, notice how this handler updates the state of the server: in case a new client joins the chat, our process will continue its execution with a ClientPortMap that includes the new (nickname, port) pair, that is, it is performing an Action that updates the state of the server process.

The messageHandler, in charge of broadcasting the messages in the chat room among all the clients, is even simpler:

messageHandler :: CastHandler ClientPortMap ChatMessage
messageHandler = handler
    handler :: ActionHandler ClientPortMap ChatMessage
    handler clients msg = do
      broadcastMessage clients msg
      continue clients

It only matches messages of type ChatMessage and broadcasts them to the other clients by using the broadcastMessage function defined above. Notice that here our process continues its execution without updating the state of the server.

The server's info handlers

Perhaps in the definition of the joinChatHandler you noticed the following mysterious line of code:

void $ monitorPort sendPort

This means that we are attaching a monitoring process to every client that is connecting to the chat server. In other words, we want our server to receive a signal whenever a client disconnects (for, example, by typing ctrl + c in the terminal). In this way, we can define a handler to match those signals and take an appropriate action on client disconnection:

disconnectHandler :: ActionHandler ClientPortMap PortMonitorNotification
disconnectHandler clients (PortMonitorNotification _ spId reason) = do
  let search = M.filter (\v -> sendPortId v == spId) clients
  case (null search, reason) of
    (False, DiedDisconnect)-> do
      let (clientName, _) = M.elemAt 0 search
          clients' = M.delete clientName clients
      broadcastMessage clients' (ChatMessage Server $ clientName ++ " has left the chat ... ")
      continue clients'
    _ -> continue clients

Whenever a client disconnects it sends a message of type PortMonitorNotification which carries the id of its SendPort. This way, we can perform a search to find which client disconnected, which is used to notify the remaining clients, and finally the server’s process continues execution with a new state which does not include the (nickname, port) pair of the disconnected client.

Next, let’s see how to implement the client that will interact with our server!

The chat client

Implementing a client that can connect to a specific chat-server is even easier. Basically we have to do 5 things:

  • Create a node for our client process to reside.
  • Search our remote chat server and get its ProcessId.
  • Send a message to the remote chat server signaling that we want to join the chat.
  • Fork a separate process to log the messages coming from other clients connected to the server.
  • Fork a separate process that constantly waits for user input and that broadcasts it as a message to the other clients connected to the remote chat server.
searchChatServer :: ChatName -> ServerAddress -> Process ProcessId
searchChatServer name serverAddr = do
  let addr = EndPointAddress (BS.pack serverAddr)
      srvId = NodeId addr
  whereisRemoteAsync srvId name
  reply <- expectTimeout 1000
  case reply of
    Just (WhereIsReply _ (Just sid)) -> return sid
    _ -> searchChatServer name serverAddr

The heart of this definition is the whereisRemoteAsync function provided by the distributed-process package. It asynchronously queries a remote node for a process that is in its local registry. That is why we pass to it the id of the remote node (NodeId) and the name of the chat server room. Notice that we recursively invoke this function until we get a WhereIsReply with the id of the remote process belonging to the chat server.

With this in mind we can cover almost all the implementation:


node <- newLocalNode transport initRemoteTable
      runChatLogger node
      runProcess node $ do
        serverPid <- searchChatServer name serverAddr
        link serverPid
        logStr "Joining chat server ... "
        logStr "Please, provide your nickname ... "
        nickName <- liftIO getLine


First, we are creating the node to host our client and run its process. Then, after getting the server’s id we have done some extra things for linking our client to the remote server, and logging a string in console indicating that you are joining that chat and must provide your nickname in the standard input. Linking our client means that if the remote server unexpectedly dies, our process will be killed.

The third step is straightforward:


rp <- callChan serverPid (JoinChatMessage nickName) :: Process (ReceivePort ChatMessage)
logStr "You have joined the chat ... "


We are only sending a message to the server of type JoinChatMessage that it will handle with the joinChatHandler explained above. This means that the server will add your client to the ClientPortMap defined in the application’s types. Moreover, we are getting the ReceivePort for the server which will be useful to get the messages that are broadcast.

Finally, we need to fork two loops, namely, one for constantly receiving messages from the chat server and another to constantly wait for user input. The first one is implemented as follows:


void $ spawnLocal $ forever $ do
  msg <- receiveChan rp
  logChatMessage msg


This means that we are spawning (forking) another process which will be listening (forever!) to messages coming from the server and it will log those messages to the standard output. For this we pass the ReceivePort from the server to receiveChan which is a function that waits for messages being sent through a channel.

The second loop has the following implementation:


forever $ do
  chatInput <- liftIO getLine
  cast serverPid (ChatMessage (Client nickName) chatInput)


Which makes use of the standard input to get text from the user that will be cast to the server in the form of a ChatMessage. The function cast allows us to send a message to a remote process without expecting a reply from it, which suits the kind of messages that our chat server dispatches with the messageHandler explained above.

And ... that’s it! We have covered all the tasks the chat client needs to perform. The full code looks like this:

launchChatClient :: ServerAddress -> Host -> Int -> ChatName -> IO ()
launchChatClient serverAddr clientHost port name  = do
  mt <- createTransport clientHost (show port) (clientHost,) defaultTCPParameters
  case mt of
    Left err -> putStrLn (show err)
    Right transport -> do
      node <- newLocalNode transport initRemoteTable
      runChatLogger node
      runProcess node $ do
        serverPid <- searchChatServer name serverAddr
        link serverPid
        logStr "Joining chat server ... "
        logStr "Please, provide your nickname ... "
        nickName <- liftIO getLine
        rp <- callChan serverPid (JoinChatMessage nickName) :: Process (ReceivePort ChatMessage)
        logStr "You have joined the chat ... "
        void $ spawnLocal $ forever $ do
          msg <- receiveChan rp
          logChatMessage msg
        forever $ do
          chatInput <- liftIO getLine
          cast serverPid (ChatMessage (Client nickName) chatInput)
          liftIO $ threadDelay 500000

Final remarks

You can check the repository for this tutorial’s source code here together with a README that explains how to launch both the chat server and the client so that you can try it out. Now that you have learned the essentials of cloud Haskell, do not hesitate to fork this repo and add your experiments or further refinements!

In case you need additional documents or readings you can visit the following links:

Published on: Apr. 10, 2018

Written by:

Sebastian Pulido Gomez

Sebastian Pulido Gomez

Subscribe to our blog

Join our community and get the latest articles, tips, and insights delivered straight to your inbox. Don’t miss it – subscribe now and be part of the conversation!

We care about your data. Check out our Privacy Policy.