Sunday, June 30, 2013

BuddyChat, a Simple Example of Akka Actors with an Akka FSM

I wrote a silly chat program in Scala to demonstrate functionality provided by something called "Akka", which is available for both Scala and Java.  It lets you easily write your program so that:

  • Everything is asynchronous, which means that things can be run at the same time when possible.  If two people want to say hello, let them say hi at the same time.  Those "hello" messages will be be displayed when they arrive, in whatever order.  Being able to run things in parallel is massively important when a computer has more than one processor, or some things might have to wait to complete, such as, perhaps, searching for a Wikipedia article.
  • Asynchronous programming is safe.  With "lower-level" implementations, it's very easy to screw up, and your software, although perhaps faster, will be prone to crashing or generating erroneous results.

Without further ado, let me introduce you to BuddyChat!  It's ugly and it's silly, but it's educational.  For people who want to see it on github.com, it's publicly available here:

https://github.com/sdanzig/buddychat

And here's a sample test run:



Description


The "gist" of this is that you're participating in a chatroom.  You run BuddyChat.  BuddyChat creates the manager of the chat.  This manager will create all the participants, both automated and human.  The one human participant it creates represents you and will provide you an interface to make it speak in the chat room.  Whenever a participant speaks, the message goes to the chat manager who forwards the message on to the other participants.

There are a couple other little features I'll describe later, but that's the brunt of it.  Here's a diagram showing this:
Aside from the slightly different names, and a couple of the messages, it's exactly as described.  The arrows represent both the actual construction of the objects, and also sending messages between them.


Construction


  1. The BuddyChat object is automatically created when the application is run.
  2. The BuddyChat object builds ChatManager.
  3. ChatManager builds the three BuddyActors (the automated chat participants)
  4. ChatManager builds UserActor.
  5. UserActor builds ConsoleActor, which accepts input from you.

Messaging


  1. BuddyChat starts off ChatManager with a CreateChat message.
  2. ChatManager receives CreateChat, then constructs the participants.
  3. ChatManager starts off all participants with a Begin message, which all but UserActor ignores.
  4. UserActor starts off ConsoleActor with an EnableConsole message.
  5. ConsoleActor sends each line of text you type as a MessageFromConsole message to UserActor.
  6. UserActor will send this text in a Speak message to the ChatManager.
  7. ChatManager will record the Speak message to its history, then forward it onto the BuddyActors.
  8. In response, each BuddyActor generates and sends a new Speak message to ChatManager.
  9. ChatManager will record each Speak message to its history, then forward them to the other participants. The BuddyActors will see the new messages are not from a human and will ignore them. The UserActor prints out the message to the screen.

There are also other messages UserActor can send ChatManager:

  • KillChat - Shut down the chat application.  Generated when UserActor receives "done".
  • StopChat - ChatManager will clear its chat history and stop accepting Speak messages.  Generated when UserActor receives "stop".
  • StartChat - ChatManager will resume accepting Speak messages. Generated when UserActor receives "start".

ChatManager's Finite State Machine (FSM)


I love finite state machines.  Let me explain what it is:

Something can be in just one out of a set of states.  When in a particular state, it behaves a particular way. When a particular condition is met, it can transition to a different state.

That's it.  They make it very easy to model potentially complex software.  Just think of what your possible states are, and what it takes to get from one state to another.  I implemented ChatManager as a finite state machine.  The states it can be in are:

  • ChatOffline
  • ChatOnline

By default, ChatManager is in the ChatOffline state.  Upon receiving the CreateChat message, it transitions to the ChatOnline state.  Receiving StopChat and StartChat messages will cause ChatManager to transition to ChatOffline and ChatOnline, respectively, if not already in the target state.

Given this, there's a negligible hiccup that occurs because, in response to a CreateChat message.  ChatManager will create a UserActor, then can send it a Begin message just before transitioning to ChatOnline.  What this means is, for a very short but existent period of time, the UserActor can send a Speak message while the state is still ChatOffline, which would consequently get ignored.  Akka provides you a way to specify something to occur during a particular transition.  In this case, ChatManager sends out the Begin message on the transition from ChatOffline to ChatOffline.


Finite State Machine Data


Okay, I lied, there's one more complexity to at least Akka's version of FSM, which I used.  Akka works very cleanly if you adhere to the design and don't use anything that's "shared".  By this, I mean you're not supposed to let things write information/data to the same place at the same time, or even read the same data if it could change at any point.  The way Akka actors (which all those Actors mentioned before, plus ChatManager are) can safely communicate are through messages.  Just like the Begin message wasn't sent out until in the ChatOnline state, it's possible to also ensure that a piece of data changes at the same time the state changes.  ChatManager uses this data-handling to manage its list of chat participants, and the chat history.


The Code


The source code for BuddyChat is available at:
To start, we'll look at the first thing that does something...

The BuddyChat Object


The first line shows how, in Akka, an actor is created.  "manager" is a unique name you can use to refer to the actor later.  It's not meant to look pretty, and adheres to a number of restrictions, such as having no spaces, but I use it for display purposes in this demo so I don't have to bother with storing a more visually appealing name.  The second line is sending a basic message to ChatManager, to tell it to get things started.  It's quite possible to send just Strings as messages, such as:

manager ! "create chat"

However, by having a specific message class "CreateChat", the compiler can warn you about typos.

ChatManager


ChatManager starts off as follows:

ChatManager inherits the functionality of Akka's Actor, and it's given the FSM trait, which allows it to operate as a finite state machine.  The number of automated participants is controlled by this hard-coded constant.  ChatManager is initialized as being in the ChatOffline state, and with no users and no chat history.  Not even empty lists, which is why it's simply Uninitialized.

Akka's structure for handling messages when in a state is quite intuitive.  It follows the paradigm: "When in state A, handle messages of type 1 this way and messages of type 2 that way."  See ChatManager's logic in the ChatOffline state:

As you can see, when offline, ChatManager can handle a CreateChat message and a StartChat message.  I won't dive too much into how case classes work in Scala, but I will point out that you don't just see "case CreateChat" here.  You see "case Event(some message type, some state data)".  This is being used not only to respond to a particular incoming message, but also to read in the state data.  It's possible to also have it respond to a message type differently depending on what your data is.  In this case, we know we only want to respond to CreateChat messages when the data is Uninitialized, so we specify this.  This ensures that if we erroneously get a CreateChat message after the chat has been created, the message will be ignored, because although the message type matches CreateChat, the state data does not match Uninitialized.

Upon reception of CreateChat, ChatManager instantiates the sole UserActor, named "user", and the three BuddyActors.  The combination of the two,

user :: list

becomes the new state data upon transitioning (going to) the ChatOnline state.  CreateChat is one message, when there is no state data, that can provoke this transition.  The other is StartChat, but only if the chat participants are already created.  That stipulation is reflected by ChatData(chatters, _).  The underscore is a placeholder for the chat history, used to convey indifference to what, if any, chat history exists.  Checking the list of chatters alone is sufficient to ensure StartChat is processed only when it should be.  Upon processing a StartChat message, ChatManager will transition to the ChatOnline state, retaining the list of chatters, and creating a new, empty chat history (List[String]()).

As mentioned before, ChatManager has some logic for immediately after transitioning from offline to online, to avoid the window of time when a UserActor can send a Speak message when ChatManager is still offline (and thus being ignored):

While the automated BuddyActors ultimately ignore the Begin message, because they only send messages in response to the user anyway, the UserActor, upon receiving a Begin message, will instruct the ConsoleActor to start receiving keyboard input.  One more quirk here.  This part:

(Uninitialized, ChatData(chatters, _)) <- Some(stateData,nextStateData)

What that is doing is ensuring the Begin message is only sent out when the chat participants are first created.  The state data goes from completely uninitialized to existing state data complete with a list of chatters.  If the change in state data doesn't match that, then nothing happens during the transition.

While in the ChatOnline state, ChatManager uses this message handling logic:


In this state, ChatManager can now accept Speak messages.  Upon receiving a Speak message, ChatManager will forward the message to all chat participants except (different from) the sender, which is where the "diff" is applied.  "forward" is used to re-send the messages rather than the typical ! because forward will send the message as if it were from the same "sender".  Akka allows you to, upon receiving a message, access the sender of that message, and if ChatManager used !, it would appear that ChatManager originated the message.  This allows the message receiver to handle a message in a different way based on who sent it.

When writing BuddyChat, I initially allowed BuddyActor to respond to all incoming messages, but ultimately the problem arose where all the BuddyActors responded to other BuddyActors repeatedly and endlessly.  By only responding to messages where the sender has the name "user", the BuddyActor is assured to avoid this issue.

Note Speak does not cause a transition.  ChatManager will "stay" at its current state.  However, it uses updated state data (ChatData) which has a chat history that includes the new message.

ChatManager also can receive a StopChat method while in ChatOnline state.  This will cause ChatManager to go to "ChatOffline" state, and while the list of chatters are preserved in the new ChatData, the chat history is replaced by an empty list of messages.

When there is no case that matches the message in the handler for the particular state, the message is dealt with in the whenUnhandled block:

In either state, ChatManager should be able to handle the KillChat message, so it makes sense to receive it here.  While whenUnhandled certainly can deal with messages that are unexpected in the current state, the fall-through logic that leads messages to whenUnhandled makes it a perfect place to handle messages that are treated the same in any state.  ChatManager does not have to clean up any resources upon shutdown, so it can call context.system.shutdown to end the application immediately.  Just for demonstration's sake, ChatManager prints out the entire chat history first, summarizing who said what.  Note that when ChatManager stores text from Speak messages, it prepends the name of the actor that generated the message.

If a message is actually unexpected, there is a catch-all handler that will log the message with current state data as a warning, but otherwise do nothing.

UserActor


A UserActor is constructed by ChatManager when it receives a CreateChat message.  Upon creation, the UserActor will create a ConsoleActor.  Very soon after UserActor is created, ChatManager will enter ChatOnline state then pass it a Begin message.  UserActor is not a finite state machine.  It will respond to the same set of messages the same way no matter the circumstances.  The messages are handled by UserActor's receive method:

Upon receiving a Begin message, UserActor sends an EnableConsole message to ConsoleActor it created.  If the UserActor tried to wait for user input directly (which I initially tried to do), it would not be able to receive any further messages.  Why is this?
An actor in Akka has a message queue which is processed one message at a time.  Waiting for keyboard input is a "blocking" operation, which means that execution ceases until keyboard input is received.  Because you need to repeatedly wait for the next line of input in a loop, the Begin message handler would never exit.  It would just repeatedly end up waiting for keyboard input.
The solution is to let ConsoleActor handle it.  If ConsoleActor receives one message and then endlessly waits for user input, this is okay, because it's running in another "thread of execution".

UserActor, after enabling the console input, will wait for an incoming MessageFromConsole.  If the text encapsulated by this message is one of the following, there is special handling:

  • "done" - Upon receiving this, UserActor will send ChatManager a KillChat message to shut down the chat system.
  • "stop" - UserActor will send ChatManager a StopChat message to disable the chatting and clear the chat history.
  • "start" - UserActor will send ChatManager a StartChat message to re-enable chatting.

If the text does not match any of those, UserActor will encapsulate the text in a Speak message and send it to ChatManager, allowing the user to communicate with the other chat participants.

From the ChatManager, UserActor can receive Speak messages which would have originated from other chat participants (BuddyActors) and then been forwarded by ChatManager.  Because the Speak message  was forwarded rather than resent, the sender is the actor that generated the message, not the ChatManager that directly sent it to the UserActor.  This allows the UserActor to pull out the originator's name to identify the sender of the message for display purposes (labeledText).

There's one more nifty thing to mention about this "matching" methodology in receive.  Later you'll see the declaration of the messages that are passed around between actors.  If all of the messages that an actor can receive have a "sealed trait", then whenever you are handling a message with this trait, Scala can confirm that you have handled every possible message that has this trait.  This is called "checking for completeness" in a pattern match.

ConsoleActor


ConsoleActor's sole purpose is to accept input from the keyboard and send it to the UserActor in a MessageFromConsole message.

It receives one message, EnableConsole, and then displays instructions enters the loop that accepts lines of input from the keyboard.  For each line, a MessageFromConsole message is sent to the UserActor, which ConsoleActor identifies as its "parent", since UserActor created it.  The only thing that can exit this loop, is when "done" is typed.  That fancy getLines.takeWhile is generating a "stream", which is a feature in Scala.

A stream can be iterated over just like a list, and each element is generated on the fly.  The takeWhile, upon detecting a value that doesn't meet a condition, will make the for loop think it's just reached the end of the list, instead of processing the nonconforming value from the stream.

After the loop has terminated, "done" is sent to UserActor to shut down the chat.

BuddyActor


BuddyActor's sole purpose is to respond to messages from UserActor.

testJust to make BuddyActor interchangeable with a fully functional UserActor, BuddyActor handles all the messages that a ChatManager might send to any other chat participant, such as UserActor:  Speak and Begin, although it will ignore the Begin message.  In the Speak message handler, the message is also ignored if the sender's name is not "user".  This prevents a BuddyActor from endlessly conversing with another BuddyActor.  When responding to a Speak message, BuddyActor will randomly generate one of three silly responses, including the text from the received message in the reply.  This inclusion is mainly to prove that BuddyActor is successfully receiving the forwarded message from the UserActor.

The random number generator, "rand", uses a "seed" that is affected by the current time in milliseconds along with the "path" of the actor, which must be unique amongst actors.  Without a random number seed, the random number generator would generate the same sequence of numbers.

Messages


The messages passed between actors are defined as follows:

The messages are given traits such that all possible messages an actor can receive share a common trait. If you accidentally remove the handling for a message in that set, the Scala compiler will warn you.  The "sealed" keyword means that all the possible classes that use that sealed trait are in the same file.  This allows the programmer to guarantee that all messages which use the trait are accounted for.

There are two subtleties used here while defining these traits:


  • A message class can have more than one trait.  The Speak message is a ChatParticipantSystemMessage and a ChatManagementSystemMessage.  That means, respectively, that it's one of the messages that a chatter can receive, and also one of the messages that ChatManager handles.
  • A trait can have another trait.  By saying that a ChatParticipantSystemMessage has the UserSystemMessage, you're saying that the set of messages with the UserSystemMessage trait is equal to or greater than the set of messages with the ChatParticipantSystemMessage trait.  Any message with the ChatParticipantSystemMessage trait also has the UserSystemMessage trait, so the set of user system messages are at least that set, and perhaps more.  In this case, there's one additional message that a UserActor can receive that the other chatters (the BuddyActors) can't receive.  The MessageFromConsole is from ConsoleActor.  ConsoleActor only communicates with UserActor, so this makes sense.


Conclusion


The BuddyChat system certainly is only meant to serve educational purposes.  However, it demonstrates many very useful technologies within both Akka, and Scala as well.  Programmers no longer need to fear multi-threaded programming as long as they properly use an actor system such as provided by Akka.  Akka's FSM can simplify the implementation of a complex system by grouping its behaviors by its possible states.  While the significant overhead of the actor framework is not suitable for applications requiring maximum performance (such as handling billions of tweets or time-sensitive stock ticker updates), in which case lower-level handling of parallel execution is recommended, Akka actors are amazingly easy to deal with and should be used otherwise.


Notes


I'm unsure if there's a way to do some form of completeness checking in the FSM handlers.  I'd guess not, but please let me know if there's a way.
I tried implementing the check for the name "user" as a guard in the pattern match:

case Speak(msg) if "user".equals(sender.path.name)

I had a case Speak(msg) after that to catch the other Speak messages and ignore them.  However, this disabled the completeness checking.  I saw in older versions of Scala that guards were handled improperly and this had been fixed, but perhaps the change was reverted, or, more likely, I'm doing something wrong.

3 comments:

  1. Hi,
    I have ported your code to Java here https://github.com/abhishekagarwal87/akka.
    This should help those people coming from java background and not familiar with scala.

    ReplyDelete
    Replies
    1. I tried to keep the Scala code simple and step people through it, but having a Java version is a really cool contribution! Thanks a ton.

      Delete
    2. This comment has been removed by the author.

      Delete