Module: Concurrent::Actor

=== Actor model

• Light-weighted running on thread-pool.
• Inspired by Akka and Erlang.
• Modular.

This Actor model implementation makes actors very cheap to create and discard. Thousands of actors can be created, allowing you to break the program into smaller maintainable pieces, without violating the single responsibility principle.

==== What is an actor model?

Actor-based concurrency is all the rage in some circles. Originally described in 1973, the actor model is a paradigm for creating asynchronous, concurrent objects that is becoming increasingly popular. Much has changed since actors were first written about four decades ago, which has led to a serious fragmentation within the actor community. There is no universally accepted, strict definition of "actor" and actor implementations differ widely between languages and libraries.

Wiki definition is pretty good: The actor model in computer science is a mathematical model of concurrent computation that treats actors as the universal primitives of concurrent digital computation: in response to a message that it receives, an actor can make local decisions, create more actors, send more messages, and determine how to respond to the next message received.

==== Why?

Concurrency is hard to get right, actors are one of many ways how to simplify the problem.

==== Quick example

An example:

class Counter < {Concurrent::Actor::Context}
  # Include context of an actor which gives this class access to reference
  # and other information about the actor

  # use initialize as you wish
  def initialize(initial_value)
    @count = initial_value
  end

  # override on_message to define actor's behaviour
  def on_message(message)
    if Integer === message
      @count += message
    end
  end
end #

=== Create new actor naming the instance 'first'.
=== Return value is a reference to the actor, the actual actor is never returned.
counter = {Counter.spawn}(:first, 5)

=== Tell a message and forget returning self.
counter.tell(1)
counter << 1
=== (First counter now contains 7.)

=== Send a messages asking for a result.
counter.ask(0).class
counter.ask(0).value

class Adder < {Concurrent::Actor::RestartingContext}
  def initialize(init)
    @count = init
  end

  def on_message(message)
    case message
    when :add
      @count += 1
    else
      # pass to ErrorsOnUnknownMessage behaviour, which will just fail
      pass
    end
  end
end 

=== <code>link: true</code> makes the actor linked to root actor and supervised
=== which is default behavior
adder = {Adder.spawn}(name: :adder, link: true, args: [1])
    # => #<Concurrent::Actor::Reference:0x7fbedd8e3d40 /adder (Adder)>
adder.parent
    # => #<Concurrent::Actor::Reference:0x7fbedbaa1e90 / (Concurrent::Actor::Root)>

=== tell and forget
adder.tell(:add).tell(:add)
    # => #<Concurrent::Actor::Reference:0x7fbedd8e3d40 /adder (Adder)>
=== ask to get result
adder.ask!(:add)                                   # => 4
=== fail the actor
adder.ask!(:bad) rescue $!
    # => #<Concurrent::Actor::UnknownMessage: :bad from #<Thread:0x007fbedb8809b8>>
=== actor is restarted with initial values
adder.ask!(:add)                                   # => 2
adder.ask!(:terminate!)                            # => true

==== Spawning actors

• .spawn and .spawn!
• AbstractContext.spawn and AbstractContext.spawn!

==== Sending messages

• Reference#tell Sends the message asynchronously to the actor and immediately returns self (the reference) allowing to chain message telling.
• Reference#ask
• Reference#ask! Sends the message synchronously and blocks until the message is processed. Raises on error.

Messages are processed in same order as they are sent by a sender. It may interleaved with messages from other senders though.

===== Immutability

Messages sent between actors should be immutable. Gems like

• Algebrick - Typed struct on steroids based on algebraic types and pattern matching
• Hamster - Efficient, Immutable, Thread-Safe Collection classes for Ruby

are very helpful.

require 'algebrick'                                # => true

=== Actor message protocol definition with Algebrick
Protocol = {Algebrick.type} do
  variants Add      = type { fields! a: Numeric, b: Numeric },
           Subtract = type { fields! a: Numeric, b: Numeric }
end                                                # => Protocol(Add | Subtract)

class Calculator < {Concurrent::Actor::RestartingContext}
  include Algebrick::Matching

  def on_message(message)
    # pattern matching on the message with deconstruction
    # ~ marks values which are passed to the block
    match message,
          (on Add.(~any, ~any) do |a, b|
            a + b
          end),
          # or using multi-assignment
          (on ~Subtract do |(a, b)|
            a - b
          end)
  end
end 

calculator = {Calculator.spawn}('calculator')
    # => #<Concurrent::Actor::Reference:0x7fbedba52d90 /calculator (Calculator)>
addition = calculator.ask Add[1, 2]
    # => <#Concurrent::Promises::Future:0x7fbedc05f7b0 pending>
subtraction = calculator.ask Subtract[1, 0.5]
    # => <#Concurrent::Promises::Future:0x7fbedd891388 pending>
results = (addition & subtraction)
    # => <#Concurrent::Promises::Future:0x7fbedc04eeb0 pending>
results.value!                                     # => [3, 0.5]

calculator.ask! :terminate!                        # => true

==== Actor definition

New actor is defined by subclassing RestartingContext, Context and defining its abstract methods. AbstractContext can be subclassed directly to implement more specific behaviour see Root implementation.

• Context

  > Basic Context of an Actor. It supports only linking and it simply terminates on error. Uses Behaviour.basic_behaviour_definition:

• RestartingContext.

  > Context of an Actor for robust systems. It supports supervision, linking, pauses on error. Uses Behaviour.restarting_behaviour_definition

Example of ac actor definition:

Message = {Struct.new} :action, :value 

class AnActor < {Concurrent::Actor::RestartingContext}
  def initialize(init)
    @counter = init
  end

  # override #on_message to define actor's behaviour on message received
  def on_message(message)
    case message.action
    when :add
      @counter = @counter + message.value
    when :subtract
      @counter = @counter - message.value
    when :value
      @counter
    else
      pass
    end
  end

  # set counter to zero when there is an error
  def on_event(event)
    if event == :reset
      @counter = 0 # ignore initial value
    end
  end
end 

an_actor = {AnActor.spawn} name: 'an_actor', args: 10 
an_actor << {Message.new}(:add, 1) << {Message.new}(:subtract, 2) 
an_actor.ask!(Message.new(:value, nil))            # => 9
an_actor << :boo << {Message.new}(:add, 1) 
an_actor.ask!(Message.new(:value, nil))            # => 1
an_actor << :terminate!
    # => #<Concurrent::Actor::Reference:0x7fbedc137688 /an_actor (AnActor)>

See methods of AbstractContext what else can be tweaked, e.g AbstractContext#default_reference_class

==== Reference

Reference is public interface of Actor instances. It is used for sending messages and can be freely passed around the application. It also provides some basic information about the actor, see PublicDelegations.

AdHoc.spawn('printer') { -> message { puts message } }
# => #<Concurrent::Actor::Reference:0x7fd0d2883218 /printer (Concurrent::Actor::Utils::AdHoc)>
#                                   ^object_id     ^path     ^context class

==== Garbage collection

Spawned actor cannot be garbage-collected until it's terminated. There is a reference held in the parent actor.

==== Parent-child relationship, name, and path

• Core#name The name of actor instance, it should be uniq (not enforced). Allows easier orientation between actor instances.
• Core#path Path of this actor. It is used for easier orientation and logging. Path is constructed recursively with: parent.path + self.name up to a .root, e.g. /an_actor/its_child.
• Core#parent A parent Actor. When actor is spawned the .current becomes its parent. When actor is spawned from a thread outside of an actor (.current is nil) .root is assigned.

==== Behaviour

Actors have modular architecture, which is achieved by combining a light core with chain of behaviours. Each message or internal event propagates through the chain allowing the behaviours react based on their responsibility.

• Actor::Behaviour::Linking:

  > Links the actor to other actors and sends actor’s events to them, like: :terminated, :paused, :resumed, errors, etc. Linked actor needs to handle those messages.

  listener = AdHoc.spawn name: :listener do
    lambda do |message|
      case message
      when Reference
        if message.ask!(:linked?)
          message << :unlink
        else
          message << :link
        end
      else
        puts "got event #{message.inspect} from #{envelope.sender}"
      end
    end
  end
  
  an_actor = AdHoc.spawn name: :an_actor, supervise: true, behaviour_definition: Behaviour.restarting_behaviour_definition do
    lambda { |message| raise 'failed'}
  end
  
  # link the actor
  listener.ask(an_actor).wait
  an_actor.ask(:fail).wait
  # unlink the actor
  listener.ask(an_actor).wait
  an_actor.ask(:fail).wait
  an_actor << :terminate!

  produces only two events, other events happened after unlinking

  got event #<RuntimeError: failed> from #<Concurrent::Actor::Reference /an_actor (Concurrent::Actor::Utils::AdHoc)>
  got event :reset from #<Concurrent::Actor::Reference /an_actor (Concurrent::Actor::Utils::AdHoc)>

• Actor::Behaviour::Awaits:

  > Accepts :await messages. Which allows to wait on Actor to process all previously send messages.

  actor << :a << :b
  actor.ask(:await).wait # blocks until :a and :b are processed

• Actor::Behaviour::Pausing:

  > Allows to pause actors on errors. When paused all arriving messages are collected and processed after the actor is resumed or reset. Resume will simply continue with next message. Reset also reinitialized context.

• Actor::Behaviour::Supervising:

  > Handles supervised actors. Handle configures what to do with failed child: :terminate!, :resume!, :reset!, or :restart!. Strategy sets :one_for_one (restarts just failed actor) or :one_for_all (restarts all child actors).

• Actor::Behaviour::Supervising:

  > Handles supervised actors. Handle configures what to do with failed child: :terminate!, :resume!, :reset!, or :restart!. Strategy sets :one_for_one (restarts just failed actor) or :one_for_all (restarts all child actors).

• Actor::Behaviour::ExecutesContext:

  > Delegates messages and events to AbstractContext instance.

• Actor::Behaviour::ErrorsOnUnknownMessage:

  > Simply fails when message arrives here. It’s usually the last behaviour.

• Actor::Behaviour::Termination:

  > Handles actor termination. Waits until all its children are terminated, can be configured on behaviour initialization.

• Actor::Behaviour::RemovesChild:

  > Removes terminated children.

If needed new behaviours can be added, or old one removed to get required behaviour.

• Context uses Array of behaviours and their construction parameters.

  [[Behaviour::SetResults, :terminate!],
   [Behaviour::RemovesChild],
   [Behaviour::Termination],
   [Behaviour::Linking],
   [Behaviour::Awaits],
   [Behaviour::ExecutesContext],
   [Behaviour::ErrorsOnUnknownMessage]]

• RestartingContext uses Array of behaviours and their construction parameters.

  [[Behaviour::SetResults, :pause!],
   [Behaviour::RemovesChild],
   [Behaviour::Termination],
   [Behaviour::Linking],
   [Behaviour::Pausing],
   [Behaviour::Supervising, :reset!, :one_for_one],
   [Behaviour::Awaits],
   [Behaviour::ExecutesContext],
   [Behaviour::ErrorsOnUnknownMessage]]

==== IO cooperation

Actors are running on shared thread poll which allows user to create many actors cheaply. Downside is that these actors cannot be directly used to do IO or other blocking operations. Blocking operations could starve the global_fast_executor. However there are two options:

• Create an regular actor which will schedule blocking operations in global_io_executor (which is intended for blocking operations) sending results back to self in messages.
• Create an actor using global_io_executor instead of global_fast_executor, e.g. AnIOActor.spawn name: :blocking, executor: Concurrent.global_io_executor.

===== Example

require 'concurrent'                               # => false

=== {Concurrent.use_simple_logger}(:WARN, STDOUT)

=== First option is to use operation pool

class ActorDoingIO < {Concurrent::Actor::RestartingContext}
  def on_message(message)
    # do IO operation
  end

  def default_executor
    Concurrent.global_io_executor
  end
end 

actor_doing_io = {ActorDoingIO.spawn} :actor_doing_io
    # => #<Concurrent::Actor::Reference:0x7fbedc146e80 /actor_doing_io (ActorDoingIO)>
actor_doing_io.executor == {Concurrent.global_io_executor}
    # => true

=== It can be also built into a pool so there is not too many IO operations

class IOWorker < {Concurrent::Actor::Context}
  def on_message(io_job)
    # do IO work
    sleep 0.1
    puts "#{path} second:#{(Time.now.to_f*100).floor} message:#{io_job}"
  end

  def default_executor
    Concurrent.global_io_executor
  end
end 

pool = {Concurrent::Actor::Utils::Pool}.spawn('pool', 2) do |index|
  IOWorker.spawn(name: "worker-#{index}")
end
    # => #<Concurrent::Actor::Reference:0x7fbedba83378 /pool (Concurrent::Actor::Utils::Pool)>

pool << 1 << 2 << 3 << 4 << 5 << 6
    # => #<Concurrent::Actor::Reference:0x7fbedba83378 /pool (Concurrent::Actor::Utils::Pool)>

=== prints two lines each second
=== /pool/worker-0 second:1414677666 message:1
=== /pool/worker-1 second:1414677666 message:2
=== /pool/worker-0 second:1414677667 message:3
=== /pool/worker-1 second:1414677667 message:4
=== /pool/worker-0 second:1414677668 message:5
=== /pool/worker-1 second:1414677668 message:6

sleep 1                                            # => 1

==== Dead letter routing

see AbstractContext#dead_letter_routing description:

Defines an actor responsible for dead letters. Any rejected message send with Reference#tell is sent there, a message with future is considered already monitored for failures. Default behaviour is to use AbstractContext#dead_letter_routing of the parent, so if no AbstractContext#dead_letter_routing method is overridden in parent-chain the message ends up in Actor.root.dead_letter_routing agent which will log warning.

==== FAQ

===== What happens if I try to supervise using a normal Context?

Alleged supervisor will receive errors from its supervised actors. They'll have to be handled manually.

===== How to change supervision strategy?

Use option behaviour_definition: {Behaviour.restarting_behaviour_definition}(:resume!) or behaviour_definition: {Behaviour.restarting_behaviour_definition}(:reset!, :one_for_all)

===== How to change behaviors?

Any existing behavior can be subclassed

===== How to implement custom restarting?

By subclassing Actor::Behaviour::Pausing and overriding Behaviour::Pausing#restart!. Implementing AbstractContext#on_event could be also considered.

We'll be happy to answer any other questions, just open an Issue or find us on https://gitter.im/ruby-concurrency/concurrent-ruby.

==== Speed

Simple benchmark Actor vs Celluloid, the numbers are looking good but you know how it is with benchmarks. Source code is in examples/actor/celluloid_benchmark.rb. It sends numbers between x actors and adding 1 until certain limit is reached.

Benchmark legend:

• mes. - number of messages send between the actors
• act. - number of actors exchanging the messages
• impl. - which gem is used

===== JRUBY

Rehearsal ---------------------------------------------------------
50000    2 concurrent  26.140000   0.610000  26.750000 (  7.761000)
50000    2 celluloid   41.440000   5.270000  46.710000 ( 17.535000)
50000  500 concurrent  11.340000   0.180000  11.520000 (  3.498000)
50000  500 celluloid   19.310000  10.680000  29.990000 ( 14.619000)
50000 1000 concurrent  10.640000   0.180000  10.820000 (  3.563000)
50000 1000 celluloid   17.840000  19.850000  37.690000 ( 18.892000)
50000 1500 concurrent  14.120000   0.290000  14.410000 (  4.618000)
50000 1500 celluloid   19.060000  28.920000  47.980000 ( 25.185000)
---------------------------------------------- total: 225.870000sec

 mes. act.      impl.       user     system      total        real
50000    2 concurrent   7.320000   0.530000   7.850000 (  3.637000)
50000    2 celluloid   13.780000   4.730000  18.510000 ( 10.756000)
50000  500 concurrent   9.270000   0.140000   9.410000 (  3.020000)
50000  500 celluloid   16.540000  10.920000  27.460000 ( 14.308000)
50000 1000 concurrent   9.970000   0.160000  10.130000 (  3.445000)
50000 1000 celluloid   15.930000  20.840000  36.770000 ( 18.272000)
50000 1500 concurrent  11.580000   0.240000  11.820000 (  3.723000)
50000 1500 celluloid   19.440000  29.060000  48.500000 ( 25.227000) (1)

===== MRI 2.1.0

Rehearsal ---------------------------------------------------------
50000    2 concurrent   4.180000   0.080000   4.260000 (  4.269435)
50000    2 celluloid    7.740000   3.100000  10.840000 ( 10.043875)
50000  500 concurrent   5.900000   1.310000   7.210000 (  6.565067)
50000  500 celluloid   12.820000   5.810000  18.630000 ( 17.320765)
50000 1000 concurrent   6.080000   1.640000   7.720000 (  6.931294)
50000 1000 celluloid   17.130000   8.320000  25.450000 ( 23.786146)
50000 1500 concurrent   6.940000   2.030000   8.970000 (  7.927330)
50000 1500 celluloid   20.980000  12.040000  33.020000 ( 30.849578)
---------------------------------------------- total: 116.100000sec

 mes. act.      impl.       user     system      total        real
50000    2 concurrent   3.730000   0.100000   3.830000 (  3.822688)
50000    2 celluloid    7.900000   2.910000  10.810000 (  9.924014)
50000  500 concurrent   5.420000   1.230000   6.650000 (  6.025579)
50000  500 celluloid   12.720000   5.540000  18.260000 ( 16.889517)
50000 1000 concurrent   5.420000   0.910000   6.330000 (  5.896689)
50000 1000 celluloid   16.090000   8.040000  24.130000 ( 22.347102)
50000 1500 concurrent   5.580000   0.760000   6.340000 (  6.038535)
50000 1500 celluloid   20.000000  11.680000  31.680000 ( 29.590774) (1)

Note (1): Celluloid is using thread per actor so this bench is creating about 1500 native threads. Actor is using constant number of threads.