Actor Model — Common Concurrency… | OneNoughtOne

0/105

Operating SystemsCommon Concurrency Patterns

Common Concurrency Patterns

LevelIntermediate

Duration90 mins

TopicCommon Concurrency Patterns

5 / 5

Actor Model

A Different Vision of Concurrency

In 1973, Carl Hewitt proposed a radically different approach to concurrent computation. Rather than threads sharing memory and coordinating through locks, he envisioned a world of independent actors —self-contained units of computation that interact only by sending messages to each other. No shared state. No locks. No race conditions. This is the actor model, and while it remained largely academic for decades, it has found powerful expression in modern systems: Erlang's fault-tolerant telecommunications systems, Akka's distributed JVM applications, and Orleans' virtual actors in cloud gaming. WhatsApp's 900+ million users are served by Erlang actors. These aren't toy systems—they're production infrastructure handling millions of concurrent connections.

What You Will Learn

By the end of this page, you will understand the actor model deeply: its core principles and how they eliminate traditional concurrency hazards, the message-passing semantics that govern actor communication, how actors enable both local and distributed concurrency with the same programming model, supervision hierarchies for fault tolerance, and practical implementation patterns across actor frameworks.

Core Principles of the Actor Model

The actor model is built on a small set of powerful principles that fundamentally change how we reason about concurrent systems. What is an Actor? An actor is the fundamental unit of computation. Each actor encapsulates: State : Private data that no other actor can directly access Behavior : Logic that processes incoming messages Mailbox : A queue of pending messages to be processed Actors are reactive —they only execute in response to messages. Between messages, an actor is dormant, consuming no CPU cycles.

The Three Actor Axioms
Axiom	Description	Implication
Create	An actor can create new actors	Enables dynamic system structure
Send	An actor can send messages to other actors	All communication is asynchronous
Become	An actor can designate behavior for next message	State transitions are explicit

Key Properties: 1. Isolation (Share Nothing) Actors share no mutable state. Each actor's memory is completely private. The only way to affect another actor's state is to send it a message—which the receiver processes when ready. This eliminates: Race conditions (no concurrent access to shared data) Deadlocks (no locks to hold and wait on) Data corruption (no torn reads/writes) 2. Asynchronous Message Passing Message sends are fire-and-forget —the sender doesn't wait for a response. Messages are queued in the receiver's mailbox and processed one at a time. This enables: Natural concurrency (send continues immediately) Location transparency (local and remote actors look the same) Non-blocking operation (no thread tied up waiting) 3. Sequential Message Processing Each actor processes one message at a time, in order of arrival. This provides: Simplified reasoning (no internal concurrency) Atomicity (message processing is atomic from actor's perspective) Consistency (state changes are sequential)

Actors vs Threads

Threads share memory and require explicit synchronization. Actors share nothing and communicate through messages. A single thread can multiplex many actors (millions in Erlang). Actor systems are typically built on a small thread pool, with runtime scheduling actors onto threads as messages arrive.

Converting Mermaid diagram...

Message Passing Semantics

Understanding the semantics of message passing is crucial for building correct actor systems. Message Delivery Guarantees: Different actor systems provide different guarantees about message delivery:

Message Delivery Semantics
Guarantee	Meaning	Examples
At-most-once	Message delivered 0 or 1 time	Default in Akka, Erlang (local)
At-least-once	Message delivered 1+ times	With acknowledgments and retry
Exactly-once	Message delivered exactly 1 time	Requires distributed transactions

Message Ordering: A critical question: if actor A sends messages M1, M2, M3 to actor B, in what order does B receive them? Guarantee : Messages from one actor to another are delivered in send order (FIFO per sender-receiver pair). Non-guarantee : Messages from different senders may interleave arbitrarily.

message_ordering.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
// Message ordering example (Akka/Scala)
 
// Actor A sends to C: M1, M2, M3
// Actor B sends to C: X1, X2
 
// GUARANTEED: C receives M1 before M2 before M3
// GUARANTEED: C receives X1 before X2
// NOT GUARANTEED: relative order of M* and X*
 
// Possible orderings C might see:
// M1, M2, M3, X1, X2    (all A first)
// M1, X1, M2, X2, M3    (interleaved)
// X1, X2, M1, M2, M3    (all B first)
// M1, X1, X2, M2, M3    (interleaved differently)
 
// For global ordering, use techniques like:
// 1. Single coordinator actor
// 2. Sequence numbers
// 3. Vector clocks

Request-Response Pattern: While sends are asynchronous, we often need responses. The actor model handles this with explicit reply messages:

request_response.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
// Akka request-response pattern
 
import akka.actor.{Actor, ActorRef, Props}
 
// Request message includes reply-to address
case class Request(query: String, replyTo: ActorRef)
case class Response(result: String)
 
class ServerActor extends Actor {
  def receive = {
    case Request(query, replyTo) =>
      val result = processQuery(query)
      replyTo ! Response(result)  // Send reply
  }
  
  def processQuery(q: String): String = s"Result for: $q"
}
 
class ClientActor(server: ActorRef) extends Actor {
  def receive = {
    case "start" =>
      // Include self as reply address
      server ! Request("hello", self)
    
    case Response(result) =>
      println(s"Got response: $result")
  }
}
 
// Modern Akka also supports ask pattern (Futures)
import akka.pattern.ask
import scala.concurrent.duration._
 
implicit val timeout = Timeout(5.seconds)
val future = server ? Request("hello", _)  // Returns Future[Response]

Tell, Don't Ask

The actor model philosophy prefers 'tell' (fire-and-forget) over 'ask' (request-response). Ask creates a Future that ties up resources waiting. Tell allows the sender to continue immediately. Design actors to tell each other what to do, not ask and wait for answers.

Erlang/OTP: The Original Actor System

Erlang was designed from the ground up around the actor model (though Erlang calls actors "processes"). Developed at Ericsson for telecommunications, Erlang pioneered many patterns now common in actor systems. Basic Erlang Actors (Processes):

erlang_basic.erl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
-module(counter).
-export([start/0, increment/1, get/1, loop/1]).
 
%% Start a new counter actor
start() ->
    spawn(fun() -> loop(0) end).  %% Create actor with initial state 0
 
%% Send increment message
increment(Pid) ->
    Pid ! increment.  %% Fire-and-forget send
 
%% Send get request (include reply address)
get(Pid) ->
    Pid ! {get, self()},
    receive
        {count, Value} -> Value
    after 5000 ->
        timeout
    end.
 
%% Actor's message processing loop
loop(Count) ->
    receive
        increment ->
            loop(Count + 1);  %% Tail recursion with new state
        
        {get, From} ->
            From ! {count, Count},  %% Send reply
            loop(Count);  %% Continue with same state
        
        stop ->
            ok  %% Exit the loop (actor terminates)
    end.
 
%% Usage:
%% Counter = counter:start().
%% counter:increment(Counter).
%% counter:increment(Counter).
%% counter:get(Counter).  %% Returns 2

Key Erlang/Actor Features: 1. Lightweight Processes Erlang processes are extremely lightweight—a process uses only ~300 bytes initially. A single Erlang VM can run millions of processes. This enables "one process per connection" patterns that would be infeasible with OS threads. 2. Immutable Data Erlang variables are immutable—once bound, they cannot change. This eliminates aliasing problems and makes message passing safe (no need to copy data). 3. Pattern Matching The receive block uses pattern matching to select which messages to handle. Unmatched messages remain in the mailbox (selective receive). 4. Let It Crash Rather than defensive programming with extensive error handling, Erlang embraces failure. Let processes crash; supervisors restart them. This leads to simpler code and more resilient systems.

WhatsApp's Scale

WhatsApp serves 900+ million users with fewer than 50 engineers, running on Erlang. A single WhatsApp server handles 2+ million concurrent connections—each connection is an Erlang process. This demonstrates the actor model's power for massive concurrency.

Akka: Actors on the JVM

Akka brings the actor model to the JVM ecosystem (Scala and Java), adding features for distributed systems, persistence, and cluster management. Akka Actor Definition:

akka_actors.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
import akka.actor.{Actor, ActorRef, ActorSystem, Props}
 
// Message definitions (case classes for pattern matching)
sealed trait CounterMessage
case object Increment extends CounterMessage
case object Decrement extends CounterMessage
case class GetCount(replyTo: ActorRef) extends CounterMessage
case class CountResult(value: Int)
 
// Actor implementation
class CounterActor extends Actor {
  // Private mutable state
  private var count: Int = 0
  
  // Message handler
  def receive: Receive = {
    case Increment =>
      count += 1
      println(s"Incremented to $count")
    
    case Decrement =>
      count -= 1
      println(s"Decremented to $count")
    
    case GetCount(replyTo) =>
      replyTo ! CountResult(count)
  }
}
 
// Usage
object CounterApp extends App {
  // Create actor system (container for actors)
  val system = ActorSystem("counter-system")
  
  // Create counter actor
  val counter = system.actorOf(Props[CounterActor], "counter")
  
  // Send messages (fire-and-forget)
  counter ! Increment
  counter ! Increment
  counter ! Decrement
  
  // For request-response, create a temporary actor or use ask pattern
  import akka.pattern.ask
  import akka.util.Timeout
  import scala.concurrent.duration._
  import scala.concurrent.ExecutionContext.Implicits.global
  
  implicit val timeout = Timeout(5.seconds)
  
  (counter ? GetCount(null)).mapTo[CountResult].foreach { result =>
    println(s"Final count: ${result.value}")
    system.terminate()
  }
}

Akka's Typed Actors: Akka Typed (now the recommended API) provides compile-time type safety for actor messages:

akka_typed.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
import akka.actor.typed.{ActorRef, ActorSystem, Behavior}
import akka.actor.typed.scaladsl.Behaviors
 
// Typed message protocol
sealed trait Command
case object Increment extends Command
case class GetCount(replyTo: ActorRef[Int]) extends Command
 
// Typed actor behavior
object Counter {
  def apply(): Behavior[Command] = counter(0)
  
  private def counter(count: Int): Behavior[Command] =
    Behaviors.receive { (context, message) =>
      message match {
        case Increment =>
          context.log.info(s"Incremented to ${count + 1}")
          counter(count + 1)  // Return new behavior with updated state
        
        case GetCount(replyTo) =>
          replyTo ! count
          Behaviors.same  // Keep same behavior
      }
    }
}
 
// Usage
@main def run(): Unit = {
  val system: ActorSystem[Command] = ActorSystem(Counter(), "typed-counter")
  
  system ! Increment
  system ! Increment
  
  // Ask pattern with typed ActorRef
  import akka.actor.typed.scaladsl.AskPattern._
  import scala.concurrent.duration._
  
  given timeout: akka.util.Timeout = 3.seconds
  given scheduler: akka.actor.typed.Scheduler = system.scheduler
  given ec: scala.concurrent.ExecutionContext = system.executionContext
  
  val result = system.ask[Int](ref => GetCount(ref))
  result.foreach(count => println(s"Count: $count"))
}

Behavior as State

In Akka Typed, state changes are expressed by returning a new Behavior. The counter(count + 1) call returns a behavior that will handle the next message with the incremented count. This functional approach makes state management explicit and safer.

Supervision and Fault Tolerance

One of the actor model's most powerful features is its approach to fault tolerance through supervision hierarchies . Rather than handling every possible error locally, actors delegate failure handling to supervisors. The Supervision Hierarchy: Actors form a tree structure: Every actor has exactly one supervisor (its parent) Parents supervise their children The root of the tree is the actor system's guardian When a child actor fails (throws an exception), the supervisor decides how to respond:

Supervision Strategies
Strategy	Action	Use Case
Resume	Continue processing with current state	Transient error, state is valid
Restart	Stop actor, create fresh instance	Corrupted state, need clean slate
Stop	Terminate the actor permanently	Unrecoverable error, give up
Escalate	Pass failure to own supervisor	Don't know how to handle

Converting Mermaid diagram...

supervision.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
import akka.actor.typed.{ActorRef, Behavior, SupervisorStrategy}
import akka.actor.typed.scaladsl.Behaviors
import scala.concurrent.duration._
 
// Worker that might fail
object Worker {
  sealed trait Command
  case class Process(data: String) extends Command
  
  def apply(): Behavior[Command] = Behaviors.receive { (context, message) =>
    message match {
      case Process(data) =>
        if (data == "fail") throw new RuntimeException("Simulated failure!")
        context.log.info(s"Processed: $data")
        Behaviors.same
    }
  }
}
 
// Supervisor with restart strategy
object Supervisor {
  sealed trait Command
  case class ProcessWork(data: String) extends Command
  
  def apply(): Behavior[Command] = Behaviors.setup { context =>
    // Create supervised worker with restart strategy
    val worker: ActorRef[Worker.Command] = context.spawn(
      Behaviors.supervise(Worker())
        .onFailure[RuntimeException](
          SupervisorStrategy.restart
            .withLimit(maxNrOfRetries = 3, withinTimeRange = 1.minute)
        ),
      "worker"
    )
    
    Behaviors.receiveMessage {
      case ProcessWork(data) =>
        worker ! Worker.Process(data)
        Behaviors.same
    }
  }
}
 
// Usage:
// supervisor ! ProcessWork("data1")   // Works
// supervisor ! ProcessWork("fail")    // Worker crashes, restarts automatically
// supervisor ! ProcessWork("data2")   // Works (on restarted worker)

Let It Crash Philosophy: Erlang/OTP pioneered the "let it crash" philosophy: Don't Defend Against Every Error : Complex defensive programming creates complex code with its own bugs Let Failures Happen : When something goes wrong, let the actor crash Supervisors Clean Up : Supervisors restart failed actors with fresh state Isolation Limits Blast Radius : Actor isolation ensures one failure doesn't corrupt the entire system Simplify Error Handling : Move error handling up the hierarchy to dedicated supervisors This approach leads to simpler code and more resilient systems. The key insight is that most errors are transient—a fresh restart often resolves the problem.

Nine Nines of Availability

Ericsson's AXD301 switch, built on Erlang/OTP supervision, achieved 99.9999999% uptime (31ms downtime per year). The supervision hierarchy allows the system to heal itself continuously—failed components restart while the rest continues operating.

Distribution and Location Transparency

The actor model naturally extends to distributed systems. Since actors communicate only through messages and share no state, an actor on another machine looks the same as a local actor. Location Transparency: Actor references (addresses) abstract away physical location. Sending a message to a remote actor uses the same syntax as sending to a local actor—the runtime handles serialization and network transport. Erlang Distribution: Erlang nodes can connect to form clusters. Any process can send messages to any process on any connected node:

erlang_distributed.erl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
%% Start Erlang nodes:
%% Node 1: erl -name node1@192.168.1.10 -setcookie secret
%% Node 2: erl -name node2@192.168.1.11 -setcookie secret
 
%% On Node 1: Register a service
-module(service).
-export([start/0, handle/0]).
 
start() ->
    Pid = spawn(fun handle/0),
    register(my_service, Pid).  %% Register with name
 
handle() ->
    receive
        {From, Message} ->
            io:format("Received: ~p from ~p~n", [Message, From]),
            From ! {reply, "Got it!"},
            handle()
    end.
 
%% On Node 2: Connect and send message
connect_and_send() ->
    %% Connect to Node 1
    net_adm:ping('node1@192.168.1.10'),
    
    %% Send message to registered process on remote node
    {my_service, 'node1@192.168.1.10'} ! {self(), "Hello from Node 2!"},
    
    %% Wait for reply
    receive
        {reply, Response} ->
            io:format("Got response: ~p~n", [Response])
    end.
 
%% The message send syntax is IDENTICAL for local and remote!

Akka Cluster: Akka provides sophisticated clustering with automatic node discovery, sharding, and distributed data:

akka_cluster.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
// Akka Cluster configuration (application.conf)
akka {
  actor {
    provider = cluster
  }
  remote.artery {
    canonical.hostname = "127.0.0.1"
    canonical.port = 2551
  }
  cluster {
    seed-nodes = [
      "akka://ClusterSystem@127.0.0.1:2551",
      "akka://ClusterSystem@127.0.0.1:2552"
    ]
  }
}
 
// Scala code for cluster-aware actor
import akka.cluster.typed.{Cluster, Subscribe}
import akka.cluster.ClusterEvent._
 
object ClusterListener {
  sealed trait Event
  case class MemberEventWrapper(event: MemberEvent) extends Event
  
  def apply(): Behavior[Event] = Behaviors.setup { context =>
    // Subscribe to cluster membership events
    val cluster = Cluster(context.system)
    cluster.subscriptions ! Subscribe(
      context.messageAdapter(MemberEventWrapper),
      classOf[MemberEvent]
    )
    
    Behaviors.receiveMessage {
      case MemberEventWrapper(MemberUp(member)) =>
        context.log.info(s"Member up: ${member.address}")
        Behaviors.same
      
      case MemberEventWrapper(MemberRemoved(member, _)) =>
        context.log.info(s"Member removed: ${member.address}")
        Behaviors.same
      
      case _ => Behaviors.same
    }
  }
}

Distribution Adds Complexity

While the actor model simplifies distribution, it doesn't eliminate distributed systems challenges. Network partitions, message loss, node failures, and ordering issues remain. The Eight Fallacies of Distributed Computing still apply. Actor systems provide tools to handle these, but you must design for failure.

Actor Patterns and Best Practices

Effective actor systems follow common patterns and practices.

Router/Load Balancer: Distribute work across multiple worker actors for parallelism:

router_pattern.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
import akka.actor.typed.{ActorRef, Behavior}
import akka.actor.typed.scaladsl.{Behaviors, Routers}
 
object Worker {
  case class Task(data: String)
  
  def apply(): Behavior[Task] = Behaviors.receive { (ctx, task) =>
    ctx.log.info(s"Processing: ${task.data}")
    Behaviors.same
  }
}
 
object WorkerPool {
  case class ProcessTask(data: String)
  
  def apply(): Behavior[ProcessTask] = Behaviors.setup { context =>
    // Create a pool of 5 workers with round-robin routing
    val pool: ActorRef[Worker.Task] = context.spawn(
      Routers.pool(5)(Worker()),
      "worker-pool"
    )
    
    Behaviors.receiveMessage {
      case ProcessTask(data) =>
        pool ! Worker.Task(data)  // Routed to next available worker
        Behaviors.same
    }
  }
}

Actor Best Practices

•Keep Actors Small: Each actor should do one thing. Complex behavior emerges from simple actor composition.
•Avoid Blocking: Never block inside an actor (no Thread.sleep, no blocking I/O). Use async patterns or dedicated blocking dispatchers.
•Immutable Messages: Messages should be immutable to avoid shared mutable state. In Scala, use case classes; in Java, use records.
•Prefer Tell Over Ask: Fire-and-forget (!) is more natural than request-response (?). Ask creates hidden complexity.
•Design for Failure: Assume actors will crash. Use supervision hierarchies. Don't put critical state only in actor memory.

Operating SystemsCommon Concurrency Patterns

Common Concurrency Patterns

LevelIntermediate

Duration90 mins

TopicCommon Concurrency Patterns

5 / 5

Actor Model

A Different Vision of Concurrency

In 1973, Carl Hewitt proposed a radically different approach to concurrent computation. Rather than threads sharing memory and coordinating through locks, he envisioned a world of independent actors —self-contained units of computation that interact only by sending messages to each other. No shared state. No locks. No race conditions. This is the actor model, and while it remained largely academic for decades, it has found powerful expression in modern systems: Erlang's fault-tolerant telecommunications systems, Akka's distributed JVM applications, and Orleans' virtual actors in cloud gaming. WhatsApp's 900+ million users are served by Erlang actors. These aren't toy systems—they're production infrastructure handling millions of concurrent connections.

What You Will Learn

By the end of this page, you will understand the actor model deeply: its core principles and how they eliminate traditional concurrency hazards, the message-passing semantics that govern actor communication, how actors enable both local and distributed concurrency with the same programming model, supervision hierarchies for fault tolerance, and practical implementation patterns across actor frameworks.

Core Principles of the Actor Model

The actor model is built on a small set of powerful principles that fundamentally change how we reason about concurrent systems. What is an Actor? An actor is the fundamental unit of computation. Each actor encapsulates: State : Private data that no other actor can directly access Behavior : Logic that processes incoming messages Mailbox : A queue of pending messages to be processed Actors are reactive —they only execute in response to messages. Between messages, an actor is dormant, consuming no CPU cycles.

The Three Actor Axioms
Axiom	Description	Implication
Create	An actor can create new actors	Enables dynamic system structure
Send	An actor can send messages to other actors	All communication is asynchronous
Become	An actor can designate behavior for next message	State transitions are explicit

Key Properties: 1. Isolation (Share Nothing) Actors share no mutable state. Each actor's memory is completely private. The only way to affect another actor's state is to send it a message—which the receiver processes when ready. This eliminates: Race conditions (no concurrent access to shared data) Deadlocks (no locks to hold and wait on) Data corruption (no torn reads/writes) 2. Asynchronous Message Passing Message sends are fire-and-forget —the sender doesn't wait for a response. Messages are queued in the receiver's mailbox and processed one at a time. This enables: Natural concurrency (send continues immediately) Location transparency (local and remote actors look the same) Non-blocking operation (no thread tied up waiting) 3. Sequential Message Processing Each actor processes one message at a time, in order of arrival. This provides: Simplified reasoning (no internal concurrency) Atomicity (message processing is atomic from actor's perspective) Consistency (state changes are sequential)

Actors vs Threads

Threads share memory and require explicit synchronization. Actors share nothing and communicate through messages. A single thread can multiplex many actors (millions in Erlang). Actor systems are typically built on a small thread pool, with runtime scheduling actors onto threads as messages arrive.

Converting Mermaid diagram...

Message Passing Semantics

Understanding the semantics of message passing is crucial for building correct actor systems. Message Delivery Guarantees: Different actor systems provide different guarantees about message delivery:

Message Delivery Semantics
Guarantee	Meaning	Examples
At-most-once	Message delivered 0 or 1 time	Default in Akka, Erlang (local)
At-least-once	Message delivered 1+ times	With acknowledgments and retry
Exactly-once	Message delivered exactly 1 time	Requires distributed transactions

Message Ordering: A critical question: if actor A sends messages M1, M2, M3 to actor B, in what order does B receive them? Guarantee : Messages from one actor to another are delivered in send order (FIFO per sender-receiver pair). Non-guarantee : Messages from different senders may interleave arbitrarily.

message_ordering.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
// Message ordering example (Akka/Scala)
 
// Actor A sends to C: M1, M2, M3
// Actor B sends to C: X1, X2
 
// GUARANTEED: C receives M1 before M2 before M3
// GUARANTEED: C receives X1 before X2
// NOT GUARANTEED: relative order of M* and X*
 
// Possible orderings C might see:
// M1, M2, M3, X1, X2    (all A first)
// M1, X1, M2, X2, M3    (interleaved)
// X1, X2, M1, M2, M3    (all B first)
// M1, X1, X2, M2, M3    (interleaved differently)
 
// For global ordering, use techniques like:
// 1. Single coordinator actor
// 2. Sequence numbers
// 3. Vector clocks

Request-Response Pattern: While sends are asynchronous, we often need responses. The actor model handles this with explicit reply messages:

request_response.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
// Akka request-response pattern
 
import akka.actor.{Actor, ActorRef, Props}
 
// Request message includes reply-to address
case class Request(query: String, replyTo: ActorRef)
case class Response(result: String)
 
class ServerActor extends Actor {
  def receive = {
    case Request(query, replyTo) =>
      val result = processQuery(query)
      replyTo ! Response(result)  // Send reply
  }
  
  def processQuery(q: String): String = s"Result for: $q"
}
 
class ClientActor(server: ActorRef) extends Actor {
  def receive = {
    case "start" =>
      // Include self as reply address
      server ! Request("hello", self)
    
    case Response(result) =>
      println(s"Got response: $result")
  }
}
 
// Modern Akka also supports ask pattern (Futures)
import akka.pattern.ask
import scala.concurrent.duration._
 
implicit val timeout = Timeout(5.seconds)
val future = server ? Request("hello", _)  // Returns Future[Response]

Tell, Don't Ask

The actor model philosophy prefers 'tell' (fire-and-forget) over 'ask' (request-response). Ask creates a Future that ties up resources waiting. Tell allows the sender to continue immediately. Design actors to tell each other what to do, not ask and wait for answers.

Erlang/OTP: The Original Actor System

Erlang was designed from the ground up around the actor model (though Erlang calls actors "processes"). Developed at Ericsson for telecommunications, Erlang pioneered many patterns now common in actor systems. Basic Erlang Actors (Processes):

erlang_basic.erl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
-module(counter).
-export([start/0, increment/1, get/1, loop/1]).
 
%% Start a new counter actor
start() ->
    spawn(fun() -> loop(0) end).  %% Create actor with initial state 0
 
%% Send increment message
increment(Pid) ->
    Pid ! increment.  %% Fire-and-forget send
 
%% Send get request (include reply address)
get(Pid) ->
    Pid ! {get, self()},
    receive
        {count, Value} -> Value
    after 5000 ->
        timeout
    end.
 
%% Actor's message processing loop
loop(Count) ->
    receive
        increment ->
            loop(Count + 1);  %% Tail recursion with new state
        
        {get, From} ->
            From ! {count, Count},  %% Send reply
            loop(Count);  %% Continue with same state
        
        stop ->
            ok  %% Exit the loop (actor terminates)
    end.
 
%% Usage:
%% Counter = counter:start().
%% counter:increment(Counter).
%% counter:increment(Counter).
%% counter:get(Counter).  %% Returns 2

Key Erlang/Actor Features: 1. Lightweight Processes Erlang processes are extremely lightweight—a process uses only ~300 bytes initially. A single Erlang VM can run millions of processes. This enables "one process per connection" patterns that would be infeasible with OS threads. 2. Immutable Data Erlang variables are immutable—once bound, they cannot change. This eliminates aliasing problems and makes message passing safe (no need to copy data). 3. Pattern Matching The receive block uses pattern matching to select which messages to handle. Unmatched messages remain in the mailbox (selective receive). 4. Let It Crash Rather than defensive programming with extensive error handling, Erlang embraces failure. Let processes crash; supervisors restart them. This leads to simpler code and more resilient systems.

WhatsApp's Scale

WhatsApp serves 900+ million users with fewer than 50 engineers, running on Erlang. A single WhatsApp server handles 2+ million concurrent connections—each connection is an Erlang process. This demonstrates the actor model's power for massive concurrency.

Akka: Actors on the JVM

Akka brings the actor model to the JVM ecosystem (Scala and Java), adding features for distributed systems, persistence, and cluster management. Akka Actor Definition:

akka_actors.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
import akka.actor.{Actor, ActorRef, ActorSystem, Props}
 
// Message definitions (case classes for pattern matching)
sealed trait CounterMessage
case object Increment extends CounterMessage
case object Decrement extends CounterMessage
case class GetCount(replyTo: ActorRef) extends CounterMessage
case class CountResult(value: Int)
 
// Actor implementation
class CounterActor extends Actor {
  // Private mutable state
  private var count: Int = 0
  
  // Message handler
  def receive: Receive = {
    case Increment =>
      count += 1
      println(s"Incremented to $count")
    
    case Decrement =>
      count -= 1
      println(s"Decremented to $count")
    
    case GetCount(replyTo) =>
      replyTo ! CountResult(count)
  }
}
 
// Usage
object CounterApp extends App {
  // Create actor system (container for actors)
  val system = ActorSystem("counter-system")
  
  // Create counter actor
  val counter = system.actorOf(Props[CounterActor], "counter")
  
  // Send messages (fire-and-forget)
  counter ! Increment
  counter ! Increment
  counter ! Decrement
  
  // For request-response, create a temporary actor or use ask pattern
  import akka.pattern.ask
  import akka.util.Timeout
  import scala.concurrent.duration._
  import scala.concurrent.ExecutionContext.Implicits.global
  
  implicit val timeout = Timeout(5.seconds)
  
  (counter ? GetCount(null)).mapTo[CountResult].foreach { result =>
    println(s"Final count: ${result.value}")
    system.terminate()
  }
}

Akka's Typed Actors: Akka Typed (now the recommended API) provides compile-time type safety for actor messages:

akka_typed.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
import akka.actor.typed.{ActorRef, ActorSystem, Behavior}
import akka.actor.typed.scaladsl.Behaviors
 
// Typed message protocol
sealed trait Command
case object Increment extends Command
case class GetCount(replyTo: ActorRef[Int]) extends Command
 
// Typed actor behavior
object Counter {
  def apply(): Behavior[Command] = counter(0)
  
  private def counter(count: Int): Behavior[Command] =
    Behaviors.receive { (context, message) =>
      message match {
        case Increment =>
          context.log.info(s"Incremented to ${count + 1}")
          counter(count + 1)  // Return new behavior with updated state
        
        case GetCount(replyTo) =>
          replyTo ! count
          Behaviors.same  // Keep same behavior
      }
    }
}
 
// Usage
@main def run(): Unit = {
  val system: ActorSystem[Command] = ActorSystem(Counter(), "typed-counter")
  
  system ! Increment
  system ! Increment
  
  // Ask pattern with typed ActorRef
  import akka.actor.typed.scaladsl.AskPattern._
  import scala.concurrent.duration._
  
  given timeout: akka.util.Timeout = 3.seconds
  given scheduler: akka.actor.typed.Scheduler = system.scheduler
  given ec: scala.concurrent.ExecutionContext = system.executionContext
  
  val result = system.ask[Int](ref => GetCount(ref))
  result.foreach(count => println(s"Count: $count"))
}

Behavior as State

In Akka Typed, state changes are expressed by returning a new Behavior. The counter(count + 1) call returns a behavior that will handle the next message with the incremented count. This functional approach makes state management explicit and safer.

Supervision and Fault Tolerance

One of the actor model's most powerful features is its approach to fault tolerance through supervision hierarchies . Rather than handling every possible error locally, actors delegate failure handling to supervisors. The Supervision Hierarchy: Actors form a tree structure: Every actor has exactly one supervisor (its parent) Parents supervise their children The root of the tree is the actor system's guardian When a child actor fails (throws an exception), the supervisor decides how to respond:

Supervision Strategies
Strategy	Action	Use Case
Resume	Continue processing with current state	Transient error, state is valid
Restart	Stop actor, create fresh instance	Corrupted state, need clean slate
Stop	Terminate the actor permanently	Unrecoverable error, give up
Escalate	Pass failure to own supervisor	Don't know how to handle

Converting Mermaid diagram...

supervision.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
import akka.actor.typed.{ActorRef, Behavior, SupervisorStrategy}
import akka.actor.typed.scaladsl.Behaviors
import scala.concurrent.duration._
 
// Worker that might fail
object Worker {
  sealed trait Command
  case class Process(data: String) extends Command
  
  def apply(): Behavior[Command] = Behaviors.receive { (context, message) =>
    message match {
      case Process(data) =>
        if (data == "fail") throw new RuntimeException("Simulated failure!")
        context.log.info(s"Processed: $data")
        Behaviors.same
    }
  }
}
 
// Supervisor with restart strategy
object Supervisor {
  sealed trait Command
  case class ProcessWork(data: String) extends Command
  
  def apply(): Behavior[Command] = Behaviors.setup { context =>
    // Create supervised worker with restart strategy
    val worker: ActorRef[Worker.Command] = context.spawn(
      Behaviors.supervise(Worker())
        .onFailure[RuntimeException](
          SupervisorStrategy.restart
            .withLimit(maxNrOfRetries = 3, withinTimeRange = 1.minute)
        ),
      "worker"
    )
    
    Behaviors.receiveMessage {
      case ProcessWork(data) =>
        worker ! Worker.Process(data)
        Behaviors.same
    }
  }
}
 
// Usage:
// supervisor ! ProcessWork("data1")   // Works
// supervisor ! ProcessWork("fail")    // Worker crashes, restarts automatically
// supervisor ! ProcessWork("data2")   // Works (on restarted worker)

Let It Crash Philosophy: Erlang/OTP pioneered the "let it crash" philosophy: Don't Defend Against Every Error : Complex defensive programming creates complex code with its own bugs Let Failures Happen : When something goes wrong, let the actor crash Supervisors Clean Up : Supervisors restart failed actors with fresh state Isolation Limits Blast Radius : Actor isolation ensures one failure doesn't corrupt the entire system Simplify Error Handling : Move error handling up the hierarchy to dedicated supervisors This approach leads to simpler code and more resilient systems. The key insight is that most errors are transient—a fresh restart often resolves the problem.

Nine Nines of Availability

Ericsson's AXD301 switch, built on Erlang/OTP supervision, achieved 99.9999999% uptime (31ms downtime per year). The supervision hierarchy allows the system to heal itself continuously—failed components restart while the rest continues operating.

Distribution and Location Transparency

The actor model naturally extends to distributed systems. Since actors communicate only through messages and share no state, an actor on another machine looks the same as a local actor. Location Transparency: Actor references (addresses) abstract away physical location. Sending a message to a remote actor uses the same syntax as sending to a local actor—the runtime handles serialization and network transport. Erlang Distribution: Erlang nodes can connect to form clusters. Any process can send messages to any process on any connected node:

erlang_distributed.erl
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
%% Start Erlang nodes:
%% Node 1: erl -name node1@192.168.1.10 -setcookie secret
%% Node 2: erl -name node2@192.168.1.11 -setcookie secret
 
%% On Node 1: Register a service
-module(service).
-export([start/0, handle/0]).
 
start() ->
    Pid = spawn(fun handle/0),
    register(my_service, Pid).  %% Register with name
 
handle() ->
    receive
        {From, Message} ->
            io:format("Received: ~p from ~p~n", [Message, From]),
            From ! {reply, "Got it!"},
            handle()
    end.
 
%% On Node 2: Connect and send message
connect_and_send() ->
    %% Connect to Node 1
    net_adm:ping('node1@192.168.1.10'),
    
    %% Send message to registered process on remote node
    {my_service, 'node1@192.168.1.10'} ! {self(), "Hello from Node 2!"},
    
    %% Wait for reply
    receive
        {reply, Response} ->
            io:format("Got response: ~p~n", [Response])
    end.
 
%% The message send syntax is IDENTICAL for local and remote!

Akka Cluster: Akka provides sophisticated clustering with automatic node discovery, sharding, and distributed data:

akka_cluster.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
// Akka Cluster configuration (application.conf)
akka {
  actor {
    provider = cluster
  }
  remote.artery {
    canonical.hostname = "127.0.0.1"
    canonical.port = 2551
  }
  cluster {
    seed-nodes = [
      "akka://ClusterSystem@127.0.0.1:2551",
      "akka://ClusterSystem@127.0.0.1:2552"
    ]
  }
}
 
// Scala code for cluster-aware actor
import akka.cluster.typed.{Cluster, Subscribe}
import akka.cluster.ClusterEvent._
 
object ClusterListener {
  sealed trait Event
  case class MemberEventWrapper(event: MemberEvent) extends Event
  
  def apply(): Behavior[Event] = Behaviors.setup { context =>
    // Subscribe to cluster membership events
    val cluster = Cluster(context.system)
    cluster.subscriptions ! Subscribe(
      context.messageAdapter(MemberEventWrapper),
      classOf[MemberEvent]
    )
    
    Behaviors.receiveMessage {
      case MemberEventWrapper(MemberUp(member)) =>
        context.log.info(s"Member up: ${member.address}")
        Behaviors.same
      
      case MemberEventWrapper(MemberRemoved(member, _)) =>
        context.log.info(s"Member removed: ${member.address}")
        Behaviors.same
      
      case _ => Behaviors.same
    }
  }
}

Distribution Adds Complexity

While the actor model simplifies distribution, it doesn't eliminate distributed systems challenges. Network partitions, message loss, node failures, and ordering issues remain. The Eight Fallacies of Distributed Computing still apply. Actor systems provide tools to handle these, but you must design for failure.

Actor Patterns and Best Practices

Effective actor systems follow common patterns and practices.

Router/Load Balancer: Distribute work across multiple worker actors for parallelism:

router_pattern.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
import akka.actor.typed.{ActorRef, Behavior}
import akka.actor.typed.scaladsl.{Behaviors, Routers}
 
object Worker {
  case class Task(data: String)
  
  def apply(): Behavior[Task] = Behaviors.receive { (ctx, task) =>
    ctx.log.info(s"Processing: ${task.data}")
    Behaviors.same
  }
}
 
object WorkerPool {
  case class ProcessTask(data: String)
  
  def apply(): Behavior[ProcessTask] = Behaviors.setup { context =>
    // Create a pool of 5 workers with round-robin routing
    val pool: ActorRef[Worker.Task] = context.spawn(
      Routers.pool(5)(Worker()),
      "worker-pool"
    )
    
    Behaviors.receiveMessage {
      case ProcessTask(data) =>
        pool ! Worker.Task(data)  // Routed to next available worker
        Behaviors.same
    }
  }
}

Actor Best Practices

•Keep Actors Small: Each actor should do one thing. Complex behavior emerges from simple actor composition.
•Avoid Blocking: Never block inside an actor (no Thread.sleep, no blocking I/O). Use async patterns or dedicated blocking dispatchers.
•Immutable Messages: Messages should be immutable to avoid shared mutable state. In Scala, use case classes; in Java, use records.
•Prefer Tell Over Ask: Fire-and-forget (!) is more natural than request-response (?). Ask creates hidden complexity.
•Design for Failure: Assume actors will crash. Use supervision hierarchies. Don't put critical state only in actor memory.

Actor Model

A Different Vision of Concurrency

Core Principles of the Actor Model

Message Passing Semantics

Erlang/OTP: The Original Actor System

Akka: Actors on the JVM

Supervision and Fault Tolerance

Distribution and Location Transparency

Actor Patterns and Best Practices

There’s more to understand.

Actor Model

A Different Vision of Concurrency

Core Principles of the Actor Model

Message Passing Semantics

Erlang/OTP: The Original Actor System

Akka: Actors on the JVM

Supervision and Fault Tolerance

Distribution and Location Transparency

Actor Patterns and Best Practices

There’s more to understand.