"I thought of objects being like biological cells and/or individual computers on a network, only able to communicate with messages" --Alan Kay, creator of Smalltalk, on the meaning of "object oriented programming"
Celluloid provides a simple and natural way to build fault-tolerant concurrent programs in Ruby. With Celluloid, you can build systems out of concurrent objects just as easily as you build sequential programs out of regular objects. Recommended for any developer, including novices, Celluloid should help ease your worries about building multithreaded Ruby programs.
Under the hood, Celluloid wraps regular objects in threads that talk to each other using messages. These concurrent objects are called "actors". When a caller wants another actor to execute a method, it literally sends it a message object telling it what method to execute. The receiver listens on its mailbox, gets the request, runs the method, and sends the caller the result. The receiver processes messages in its inbox one-at-a-time, which means that you don't need to worry about synchronizing access to an object's instance variables.
In addition to that, Celluloid also gives you the ability to call methods asynchronously, so the receiver to do things in the background for you without the caller having to sit around waiting for the result.
Like Celluloid? Join the Google Group
Celluloid works on Ruby 1.9.2, JRuby 1.6 (in 1.9 mode), and Rubinius 2.0. JRuby or Rubinius are the preferred platforms as they support true concurrent threads.
To use JRuby in 1.9 mode, you'll need to pass the "--1.9" command line option to the JRuby executable, or set the "JRUBY_OPTS=--1.9" environment variable.
Celluloid works on Rubinius in either 1.8 or 1.9 mode.
To use Celluloid, define a normal Ruby class that includes Celluloid:
class Sheen
include Celluloid
def initialize(name)
@name = name
end
def set_status(status)
@status = status
end
def report
"#{@name} is #{@status}"
end
end
Now when you create new instances of this class, they're actually concurrent objects, each running in their own thread:
>> charlie = Sheen.new "Charlie Sheen"
=> #<Celluloid::Actor(Sheen:0x00000100a312d0) @name="Charlie Sheen">
>> charlie.set_status "winning!"
=> "winning!"
>> charlie.report
=> "Charlie Sheen is winning!"
>> charlie.set_status! "asynchronously winning!"
=> nil
>> charlie.report
=> "Charlie Sheen is asynchronously winning!"
You can call methods on this concurrent object just like you would any other Ruby object. The Sheen#set_status method works exactly like you'd expect, returning the last expression evaluated.
However, Celluloid's secret sauce kicks in when you call banged predicate methods (i.e. methods ending in !). Even though the Sheen class has no set_status! method, you can still call it. Why is this? Because bang methods have a special meaning in Celluloid. (Note: this also means you can't define bang methods on Celluloid classes and expect them to be callable from other objects)
Adding a bang to the end of a method instructs Celluloid that you would like for the given method to be called asynchronously. This means that rather than the caller waiting for a response, the caller sends a message to the concurrent object that you'd like the given method invoked, and then the caller proceeds without waiting for a response. The concurrent object receiving the message will then process the method call in the background.
Adding a bang to a method name is a convention in Ruby used to indicate that the method is in some way "dangerous", and in Celluloid this is no exception. You have no guarantees that just because you made an asynchronous call it was ever actually invoked. Asynchronous calls will never raise an exception, even if an exception occurs when the receiver is processing it. Worse, unhandled exceptions will crash the receiver, and making an asynchronous call to a crashed object will not raise an error.
However, you can still handle errors created by asynchronous calls using two features of Celluloid called supervisors and linking. See the corresponding sections below for more information.
Futures allow you to request a computation and get the result later. There are two types of futures supported by Celluloid: method futures and block futures. Method futures work by invoking the future method on an actor. This method is analogous to the typical send method in that it takes a method name, followed by an arbitrary number of arguments, and a block. Let's invoke the report method from the charlie object used in the above example using a future:
>> future = charlie.future :report
=> #<Celluloid::Future:0x000001009759b8>
>> future.value
=> "Charlie Sheen is winning!"
The call to charlie.future immediately returns a Celluloid::Future object, regardless of how long it takes to execute the "report" method. To obtain the result of the call to "report", we call the value method of the future object. This call will block until the value returned from the method call is available (i.e. the method has finished executing). If an exception occured during the method call, the call to future.value will reraise the same exception.
Futures also allow you to background the computation of any block:
>> future = Celluloid::Future.new { 2 + 2 }
=> #<Celluloid::Future:0x000001008425f0>
>> future.value
=> 4
One thing to be aware of when using futures: always make sure to obtain the value of any future you make. Futures create a thread in the background which will continue to run until the future's value is obtained. Failing to obtain the value of futures you create will leak threads.
You may be familiar with tools like Monit or God which keep an eye on your applications and restart them when they crash. Celluloid supervisors work in a similar fashion, except instead of monitoring applications, they monitor individual actors and restart them when they crash. Crashes occur whenever an unhandled exception is raised anywhere within an actor.
To supervise an actor, start it with the supervise method. Using the Sheen class from the example above:
>> supervisor = Sheen.supervise "Charlie Sheen"
=> #<Celluloid::Supervisor(Sheen) "Charlie Sheen">
This created a new Celluloid::Supervisor actor, and also created a new Sheen actor, giving its initialize method the argument "Charlie Sheen". The supervise method has the same method signature as new. However, rather than returning the newly created actor, supervise returns the supervisor. To retrieve the actor that the supervisor is currently using, use the Celluloid::Supervisor#actor method:
>> supervisor = Sheen.supervise "Charlie Sheen"
=> #<Celluloid::Supervisor(Sheen) "Charlie Sheen">
>> charlie = supervisor.actor
=> #<Celluloid::Actor(Sheen:0x00000100a312d0)>
Supervisors can also automatically put actors into the actor registry using the supervise_as method:
>> Sheen.supervise_as :charlie, "Charlie Sheen"
=> #<Celluloid::Supervisor(Sheen) "Charlie Sheen">
>> charlie = Celluloid::Actor[:charlie]
=> #<Celluloid::Actor(Sheen:0x00000100a312d0)>
In this case, the supervisor will ensure that an actor of the Sheen class, created using the given arguments, is aways available by calling Celluloid::Actor[:charlie]. The first argument to supervise_as is the name you'd like the newly created actor to be registered under. The remaining arguments are passed to initialize just like you called new.
See the "Registry" section below for more information on the actor registry
Whenever any unhandled exceptions occur in any of the methods of an actor, that actor crashes and dies. Let's start with an example:
class JamesDean
include Celluloid
class CarInMyLaneError < StandardError; end
def drive_little_bastard
raise CarInMyLaneError, "that guy's gotta stop. he'll see us"
end
end
Now, let's have James drive Little Bastard and see what happens:
>> james = JamesDean.new
=> #<Celluloid::Actor(JamesDean:0x1068)>
>> james.drive_little_bastard!
=> nil
>> james
=> #<Celluloid::Actor(JamesDean:0x1068) dead>
When we told james asynchronously to drive Little Bastard, it killed him! If we were Elizabeth Taylor, co-star in James' latest film at the time of his death, we'd certainly want to know when he died. So how can we do that?
Actors can link to other actors they're interested in and want to receive crash notifications from. In order to receive these events, we need to use the trap_exit method to be notified of them. Let's look at how a hypothetical Elizabeth Taylor object could be notified that James Dean has crashed:
class ElizabethTaylor
include Celluloid
trap_exit :actor_died
def actor_died(actor, reason)
puts "Oh no! #{actor.inspect} has died because of a #{reason.class}"
end
end
We've now used the trap_exit method to configure a callback which is invoked whenever any linked actors crashed. Now we need to link Elizabeth to James so James' crash notifications get sent to her:
>> james = JamesDean.new
=> #<Celluloid::Actor(JamesDean:0x11b8)>
>> elizabeth = ElizabethTaylor.new
=> #<Celluloid::Actor(ElizabethTaylor:0x11f0)>
>> elizabeth.link james
=> #<Celluloid::Actor(JamesDean:0x11b8)>
>> james.drive_little_bastard!
=> nil
Oh no! #<Celluloid::Actor(JamesDean:0x11b8) dead> has died because of a JamesDean::CarInMyLaneError
Elizabeth called the link method to receive crash events from James. Because Elizabeth was linked to James, when James crashed, James' exit message was sent to her. Because Elizabeth was trapping the exit messages she received using the trap_exit method, the callback she specified was invoked, allowing her to take action (in this case, printing the error). But what would happen if she weren't trapping exits? Let's break James apart into two separate objects, one for James himself and one for Little Bastard, his car:
class PorscheSpider
include Celluloid
class CarInMyLaneError < StandardError; end
def drive_on_route_466
raise CarInMyLaneError, "head on collision :("
end
end
class JamesDean
include Celluloid
def initialize
@little_bastard = PorscheSpider.new_link
end
def drive_little_bastard
@little_bastard.drive_on_route_466
end
end
If you take a look in JamesDean#initialize, you'll notice that to create an instance of PorcheSpider, James is calling the new_link method.
This method works similarly to new, except it combines new and link into a single call.
Now what happens if we repeat the same scenario with Elizabeth Taylor watching for James Dean's crash?
>> james = JamesDean.new
=> #<Celluloid::Actor(JamesDean:0x1108) @little_bastard=#<Celluloid::Actor(PorscheSpider:0x10ec)>>
>> elizabeth = ElizabethTaylor.new
=> #<Celluloid::Actor(ElizabethTaylor:0x1144)>
>> elizabeth.link james
=> #<Celluloid::Actor(JamesDean:0x1108) @little_bastard=#<Celluloid::Actor(PorscheSpider:0x10ec)>>
>> james.drive_little_bastard!
=> nil
Oh no! #<Celluloid::Actor(JamesDean:0x1108) dead> has died because of a PorscheSpider::CarInMyLaneError
When Little Bastard crashed, it killed James as well. Little Bastard killed James, and because Elizabeth was trapping James' exit events, she received the notification of James' death.
Actors that are linked together propagate their error messages to all other actors that they're linked to. Unless those actors are trapping exit events, those actors too will die, like James did in this case. If you have many, many actors linked together in a large object graph, killing one will kill them all unless they are trapping exits.
This allows you to factor your problem into several actors. If an error occurs in any of them, it will kill off all actors used in a particular system. In general, you'll probably want to have a supervisor start a single actor which is in charge of a particular part of your system, and have that actor new_link to other actors which are part of the same system. If any error occurs in any of these actors, all of them will be killed off and the entire subsystem will be restarted by the supervisor in a clean state.
If, for any reason, you've linked to an actor and want to sever the link, there's a corresponding unlink method to remove links between actors.
Celluloid lets you register actors so you can refer to them symbolically. You can register Actors using Celluloid::Actor[]:
>> james = JamesDean.new
=> #<Celluloid::Actor(JamesDean:0x80c27ce0)>
>> Celluloid::Actor[:james] = james
=> #<Celluloid::Actor(JamesDean:0x80c27ce0)>
>> Celluloid::Actor[:james]
=> #<Celluloid::Actor(JamesDean:0x80c27ce0)>
The Celluloid::Actor constant acts as a hash, allowing you to register actors under the name of your choosing, and access actors by name rather than reference. This is important because actors may crash. If you're attempting to reference an actor explicitly by storing it in a variable, you may be holding onto a reference to a crashed copy of that actor, rather than talking to a working, freshly-restarted version.
The main use of the registry is for interfacing with actors that are automatically restarted by supervisors when they crash.
Signaling is an advanced technique similar to condition variables in typical multithreaded programming. One method within a concurrent object can suspend itself waiting for a particular event, allowing other methods to run. Another method can then signal all methods waiting for a particular event, and even send them a value in the process:
class SignalingExample
include Celluloid
attr_reader :signaled
def initialize
@signaled = false
end
def wait_for_signal
value = wait :ponycopter
@signaled = true
value
end
def send_signal(value)
signal :ponycopter, value
end
end
The wait_for_signal method in turn calls a method called "wait". Wait suspends the running method until another method of the same object calls the "signal" method with the same label.
The send_signal method of this class does just that, signaling "ponycopter" with the given value. This value is returned from the original wait call.
Celluloid provides a separate class of actors which run alongside I/O operations. These actors are slower and more heavyweight and should only be used when writing actors that also handle I/O operations.
To create an IO actor, include Celluloid::IO:
class IOActor
include Celluloid::IO
def initialize(sock)
@sock = sock
end
def read
wait_readable(@sock) do
@sock.read_nonblock
end
end
end
By default, Celluloid will log any errors and backtraces from any crashing actors to STDOUT. However, if you wish you can use any logger which is duck typed with the standard Ruby Logger API (i.e. it implements the #error method). For example, if you're using Celluloid within a Rails application, you'll probably want to do:
Celluloid.logger = Rails.logger
The logger class you specify must be thread-safe, although with a logging API about the worst you have to worry about with thread safety bugs is out-of-order messages in the log.
Celluloid is fundamentally a messaging system which uses thread-safe proxies to manage all inter-object communication in the system. While the goal of these proxies is to make it simple for you to write concurrent programs by applying the uniform access principle to thread-safe inter-object messaging, you can't simply forget they're there.
The thread-safety guarantees Celluloid provides around synchronizing access to instance variables only work so long as all access to actors go through the proxy objects. If the real objects that Celluloid is wrapping in an actor manage to leak out of the system, all hell will break loose.
Here are a few rules you can follow to keep this from happening:
-
NEVER RETURN SELF (or pass self as an argument to other actors): in cases where you want to pass an actor around to other actors or threads, use Celluloid.current_actor. If you grab the latest master of Celluloid off of Github, you can just use the #current_actor method when you are inside of an actor itself.
-
Don't mutate the state of objects you've sent in calls to other actors: This means you must think about data in one of two different ways: either you "fire and forget" the data, leaving it for other actors to do with what they will, or you must treat it as immutable if you have any plans of sharing it with other actors. If you're paranoid (and when you're dealing with concurrency, there's nothing wrong with being paranoid), you can freeze objects so you can detect subsequent mutations (or rather, turn attempts at mutation into errors).
-
Don't mix Ruby thread primitives and calls to other actors: if you make a call to another actor with a mutex held, you're doing it wrong. It's perfectly fine and strongly encouraged to call out to thread safe libraries from Celluloid actors. However, if you're using libraries that acquire mutexes and then execute callbacks (e.g. they take a block while they're holding a mutex) the guarantees that Celluloid provides will become weak and you may encounter deadlocks.
-
Use Fibers at your own risk: Celluloid employs Fibers as an intrinsic part of how it implements actors. While it's possible for certain uses of Fibers to cooperatively work alongside how Celluloid behaves, in most cases you'll be writing a check you can't afford. So please ask yourself: why are you using Fibers, and why can't it be solved by a block? If you've got a really good reason and you're feeling lucky, knock yourself out.
Ruby actually has a pretty good story when it comes to thread safety. The best strategy for thread safety is to share as little state as possible, and if you do share state, you should never mutate it. The worry of anyone stepping into a thread safe world is that you're using a bunch of legacy libraries with dubious thread safety. Who knows what those crazy library authors were doing?
Relax people. You're using a language where somebody can change what the '+' operator does to numbers. So why aren't we afraid to add numbers? Who knows what those crazy library authors may have done! Instead of freaking out, we can learn some telltale signs of things that will cause thread safety problems in Ruby programs so we can identify potential problem libraries just from how their APIs behave.
The #1 thread safety issue to look out for in a Ruby library is if it provides some sort of singleton access to a particular object through a class method, e.g MyClass.zomgobject, as opposed to asking you do do MyClass.new. If you aren't allocating the object, it isn't yours, it's somebody else's, and you better damn well make sure you can share nice, or you shouldn't play with it at all.
How do we share nicely? Let's find out by first looking at a thread-unsafe version of a singleton method:
class Foo
def self.current
@foo ||= Foo.new
end
end
Seems bad. All threads will share access to the same Foo object, and there's also a secondary bug here which means when the object is first being allocated and memoized as @foo. The first thread that tries to allocate it may get a different version than all the other threads because the memo value it set got clobbered by another thread because it's unsynchronized.
What else can we do? It depends on why the library is memoizing. Perhaps the Foo object has some kind of setup cost, such as making a network connection, and we want to keep it around instead of setting it up and tearing it down every time. If that's the case, the simplest thing we can do to make this code thread safe is to create a thread-specific memo of the object:
class Foo
def self.current
Thread.current[:foo] ||= Foo.new
end
end
Keep in mind that this will require N Foo objects for N threads. If each object is wrapping a network connection, this might be a concern. That said, if you see this pattern employed in the singleton methods of a library, it's most likely thread safe, provided that Foo doesn't do other wonky things.
- Fork Celluloid on github
- Make your changes and send me a pull request
- If I like them I'll merge them and give you commit access to my repository
Copyright (c) 2011 Tony Arcieri. See LICENSE.txt for further details.