com-lihaoyi / castor   0.3.0

GitHub

Castor is a lightweight, typed Actor library for Scala and Scala.js

Scala versions: 3.x 2.13 2.12
Scala.js versions: 1.x
Scala Native versions: 0.4

Castor 0.3.0

Join the chat at https://gitter.im/lihaoyi/castor

Castor is a lightweight, typed Actor library for Scala and Scala.js, making it very easy for you to define concurrent data pipelines or state machines.

// Mill
ivy"com.lihaoyi::castor:0.3.0"

// SBT
"com.lihaoyi" %% "castor" % "0.3.0"

// Scala.js / Scala Native
ivy"com.lihaoyi::castor::0.3.0"
"com.lihaoyi" %%% "castor" % "0.3.0"

Castor Actors are much more lightweight than a full-fledged framework like Akka: Castor does not support any sort of distribution or clustering, and runs entirely within a single process. Castor Actors are garbage collectible, and you do not need to manually terminate them or manage their lifecycle. Castor also provides tools to help test your actors deterministically - running them single threaded and waiting for async processing to complete - so you can test your actor logic without unreliable timing-based assertions.

Castor Actors can be run on both multithreaded and single-threaded environments, including compiled to Javascript via Scala.js.

Castor actors are used heavily in the Cask web framework to model websocket server and client connections, in the databricks/devbox file synchronizer, and in several other applications on both JVM and JS runtimes.

Castor Actors

At their core, Actors are simply objects who receive messages via a send method, and asynchronously process those messages one after the other:

trait Actor[T]{
  def send(t: T): Unit

  def sendAsync(f: scala.concurrent.Future[T]): Unit
}

This processing happens in the background, and can take place without blocking. After a messsage is sent, the thread or actor that called .send() can immediately go on to do other things, even if the message hasn't been processed yet. Messages sent to an actor that is already busy will be queued up until the actor is free.

Note that Actor is parametrized on the type T; T specifies what messages a particular Actor is expected to receive, and is checked at compile to to make sure your actors sending messages to one another are wired up correctly.

Castor provides three primary classes you can inherit from to define actors:

abstract class SimpleActor[T]()(implicit ac: Context) extends Actor[T]{
  def run(msg: T): Unit
}

abstract class BatchActor[T]()(implicit ac: Context) extends Actor[T]{
  def runBatch(msgs: Seq[T]): Unit
}

abstract class StateMachineActor[T]()(implicit ac: Context) extends Actor[T]() {
  class State(val run: T => State)
  protected[this] def initialState: State
}

SimpleActor works by providing a run function that will be run on each message. BatchActor allows you to provide a runBatch function that works on groups of messages at a time: this is useful when message processing can be batched together for better efficiency, e.g. making batched database queries instead of many individual. StateMachineActor allows you to define actors via a set of distinct states, each of which has a separate run callback that transitions the actor to a different state.

Note that any exception that is thrown while an Actor is processing a message (or batch of messages, in the case of BatchActor) is simply reported to the castor.Context's reportFailure function: the default just prints to the console using .printStackTrace(), but you can hook in to pass the exceptions elsewhere e.g. if you have a remote error aggregating service. The actor continues processing messages after the failure in the state that it was left in.

Castor Actors are meant to manage mutable state internal to the Actor. Note that it is up to you to mark the state private to avoid accidental external access. Each actor may run on a different thread, and the same actor may run on different threads at different times, so you should ensure you do not mutate shared mutable state otherwise you risk race conditions.

Writing Actors

To introduce you to using Castor Actors for writing concurrent data pipelines, we will explore a few examples using Castor to write an asynchronous, concurrent logging pipeline. This logging pipeline will receive logs from an application, and process them in the background without needing the application to stop and wait for it.

Example: Asynchronous Logging using an Actor

Here is a small demonstration of using a castor.SimpleActor to perform asynchronous logging to disk:

class Logger(log: os.Path, old: os.Path, rotateSize: Int)
            (implicit ac: castor.Context) extends castor.SimpleActor[String]{
  def run(s: String) = {
    val newLogSize = logSize + s.length + 1
    if (newLogSize <= rotateSize) logSize = newLogSize
    else {
      logSize = s.length
      os.move(log, old, replaceExisting = true)
    }
    os.write.append(log, s + "\n", createFolders = true)
  }
  private var logSize = 0
}

implicit val ac = new castor.Context.Test()

val logPath = os.pwd / "out" / "scratch" / "log.txt"
val oldPath  = os.pwd / "out" / "scratch" / "log-old.txt"

val logger = new Logger(logPath, oldPath, rotateSize = 50)

In the above example, we are defining a single Logger actor class, which we are instantiating once as val logger. We can now send as many messages as we want via logger.send: while the processing of a message make take some time (here are are both writing to disk, as well as providing log-rotation to avoid the logfile growing in size forever) the fact that it's in a separate actor means the processing happens in the background without slowing down the main logic of your program. Castor Actors process messages one at a time, so by putting the file write-and-rotate logic inside an Actor we can be sure to avoid race conditions that may arise due to multiple threads mangling the same file at once.

Here's the result of sending messages to the actor:

logger.send("I am cow")
logger.send("hear me moo")
logger.send("I weight twice as much as you")
logger.send("And I look good on the barbecue")
logger.send("Yoghurt curds cream cheese and butter")
logger.send("Comes from liquids from my udder")
logger.send("I am cow, I am cow")
logger.send("Hear me moo, moooo")

// Logger hasn't finished yet, running in the background
ac.waitForInactivity()
// Now logger has finished

os.read.lines(oldPath) ==> Seq("Comes from liquids from my udder")
os.read.lines(logPath) ==> Seq("I am cow, I am cow", "Hear me moo, moooo")

Using Actors is ideal for scenarios where the dataflow is one way: e.g. when logging, you only write logs, and never need to wait for the results of processing them.

All Castor actors require a castor.Context, which is an extended scala.concurrent.ExecutionContext. Here we are using Context.Test, which also provides the handy waitForInactivity() method which blocks until all asynchronous actor processing has completed.

Note that logger.send is thread-safe: multiple threads can be sending logging messages to the logger at once, and the .send method will make sure the messages are properly queued up and executed one at a time. This has the advantage that even when we need to stop the logging and rotate the log file, we do not need to worry about other messages being written to the log file while that is happening. The application sending logs to logger also does not need to stop and wait for the log file rotation to complete, and can proceed with its execution while the logger does its work in the background.

Strawman: Synchronized Logging

To illustrate further the use case of actors, let us consider the earlier example but using a synchronized method instead of a castor.SimpleActor to perform the logging:

val rotateSize = 50
val logPath = os.pwd / "out" / "scratch" / "log.txt"
val oldPath = os.pwd / "out" / "scratch" / "log-old.txt"

var logSize = 0

def logLine(s: String): Unit = synchronized{
  val newLogSize = logSize + s.length + 1
  if (newLogSize <= rotateSize) logSize = newLogSize
  else {
    logSize = 0
    os.move(logPath, oldPath, replaceExisting = true)
  }

  os.write.append(logPath, s + "\n", createFolders = true)
}

logLine("I am cow")
logLine("hear me moo")
logLine("I weight twice as much as you")
logLine("And I look good on the barbecue")
logLine("Yoghurt curds cream cheese and butter")
logLine("Comes from liquids from my udder")
logLine("I am cow, I am cow")
logLine("Hear me moo, moooo")

os.read(oldPath).trim() ==> "Yoghurt curds cream cheese and butter\nComes from liquids from my udder"
os.read(logPath).trim() ==> "I am cow, I am cow\nHear me moo, moooo"

This is similar to the earlier Actor example, but with two main caveats:

  • Your program execution stops when calling logLine, until the call to logLine completes. Thus the calls to logLine can end up slowing down your program, even though your program really doesn't need the result of logLine in order to make progress

  • Since logLine ends up managing some global mutable state (writing to and rotating log files) we need to make it synchronized. That means that if multiple threads in your program are calling logLine, it is possible that some threads will be blocked waiting for other threads to complete their logLine calls.

Using Castor Actors to perform logging avoids both these issues: calls to logger.send happen in the background without slowing down your main program, and multiple threads can call logger.send without being blocked by each other.

Parallelism using Actor Pipelines

Another advantage of Actors is that you can get pipelined parallelism when processing data. In the following example, we define two actor classes Writer and Logger, and two actors val writer and val logger. Writer handles the same writing-strings-to-disk-and-rotating-log-files logic we saw earlier, while Logger adds another step of encoding the data (here just using Base64) before it gets written to disk:

class Writer(log: os.Path, old: os.Path, rotateSize: Int)
            (implicit ac: castor.Context) extends castor.SimpleActor[String]{
  def run(s: String) = {
    val newLogSize = logSize + s.length + 1
    if (newLogSize <= rotateSize) logSize = newLogSize
    else {
      logSize = s.length
      os.move(log, old, replaceExisting = true)
    }
    os.write.append(log, s + "\n", createFolders = true)
  }
  private var logSize = 0
}

class Logger(dest: castor.Actor[String])
            (implicit ac: castor.Context) extends castor.SimpleActor[String]{
  def run(s: String) = dest.send(java.util.Base64.getEncoder.encodeToString(s.getBytes))
}

implicit val ac = new castor.Context.Test()

val logPath = os.pwd / "out" / "scratch" / "log.txt"
val oldPath  = os.pwd / "out" / "scratch" / "log-old.txt"

val writer = new Writer(logPath, oldPath, rotateSize = 50)
val logger = new Logger(writer)

Although we have added another Base64 encoding step to the logging process, this new step lives in a separate actor from the original write-to-disk step, and both of these can run in parallel as well as in parallel with the main logic. By constructing our data processing flows using Actors, we can take advantage of pipeline parallelism to distribute the processing over multiple threads and CPU cores, so adding steps to the pipeline neither slows it down nor does it slow down the execution of the main program.

We can send messages to this actor and verify that it writes lines to the log file base64 encoded:

logger.send("I am cow")
logger.send("hear me moo")
logger.send("I weight twice as much as you")
logger.send("And I look good on the barbecue")
logger.send("Yoghurt curds cream cheese and butter")
logger.send("Comes from liquids from my udder")
logger.send("I am cow, I am cow")
logger.send("Hear me moo, moooo")

ac.waitForInactivity()

os.read(oldPath) ==> "Q29tZXMgZnJvbSBsaXF1aWRzIGZyb20gbXkgdWRkZXI=\n"
os.read(logPath) ==> "SSBhbSBjb3csIEkgYW0gY293\nSGVhciBtZSBtb28sIG1vb29v\n"

def decodeFile(p: os.Path) = {
  os.read.lines(p).map(s => new String(java.util.Base64.getDecoder.decode(s)))
}

decodeFile(oldPath) ==> Seq("Comes from liquids from my udder")
decodeFile(logPath) ==> Seq("I am cow, I am cow", "Hear me moo, moooo")

You can imagine adding additional stages to this actor pipeline, to perform other sorts of processing, and have those additional stages running in parallel as well.

Batch Logging using BatchActor

Sometimes it is more efficient for an Actor to handle all incoming messages at once. You may be working with a HTTP API that lets you send one batch request rather than a hundred small ones, or with a database that lets you send one batch query to settle all incoming messages. In these situations, you can use a BatchActor.

This example again shows a logging pipeline, but instead of the two stages being "encoding" and "writing to disk", our two stages are "handling log rotating" and "batch writing":

sealed trait Msg
case class Text(value: String) extends Msg
case class Rotate() extends Msg
class Writer(log: os.Path, old: os.Path)
            (implicit ac: castor.ContextContext) extends castor.BatchActor[Msg]{
  def runBatch(msgs: Seq[Msg]): Unit = {
    msgs.lastIndexOf(Rotate()) match{
      case -1 => os.write.append(log, groupMsgs(msgs), createFolders = true)
      case rotateIndex =>
        val prevRotateIndex = msgs.lastIndexOf(Rotate(), rotateIndex - 1)
        if (prevRotateIndex != -1) os.remove.all(log)
        os.write.append(log, groupMsgs(msgs.slice(prevRotateIndex, rotateIndex)), createFolders = true)
        os.move(log, old, replaceExisting = true)
        os.write.over(log, groupMsgs(msgs.drop(rotateIndex)), createFolders = true)
    }
  }
  def groupMsgs(msgs: Seq[Msg]) = msgs.collect{case Text(value) => value}.mkString("\n") + "\n"
}

class Logger(dest: Actor[Msg], rotateSize: Int)
            (implicit ac: castor.Context) extends castor.SimpleActor[String]{
  def run(s: String) = {
    val newLogSize = logSize + s.length + 1
    if (newLogSize <= rotateSize) logSize = newLogSize
    else {
      logSize = s.length
      dest.send(Rotate())
    }
    dest.send(Text(s))
  }
  private var logSize = 0
}

implicit val ac = new castor.Context.Test()

val logPath = os.pwd / "out" / "scratch" / "log.txt"
val oldPath  = os.pwd / "out" / "scratch" / "log-old.txt"

val writer = new Writer(logPath, oldPath)
val logger = new Logger(writer, rotateSize = 50)

Here the Logger actor takes incoming log lines and decides when it needs to trigger a log rotation, while sending both the log lines and rotation commands as Text and Rotate commands to the Writer batch actor which handles batches of these messages via its runBatch method. Writer filters through the list of incoming messages to decide what it needs to do: either there are zero Rotate commands and it simply appends all incoming Texts to the log file, or there are one-or-more Rotate commands it needs to do a log rotation, writing the batched messages once to the log file pre- and post-rotation.

We can send messages to the logger and verify that it behaves the same as the SimpleActor example earlier:

logger.send("I am cow")
logger.send("hear me moo")
logger.send("I weight twice as much as you")
logger.send("And I look good on the barbecue")
logger.send("Yoghurt curds cream cheese and butter")
logger.send("Comes from liquids from my udder")
logger.send("I am cow, I am cow")
logger.send("Hear me moo, moooo")

ac.waitForInactivity()

os.read.lines(oldPath) ==> Seq("Comes from liquids from my udder")
os.read.lines(logPath) ==> Seq("I am cow, I am cow", "Hear me moo, moooo")

Using a BatchActor here helps reduce the number of writes to the filesystem: no matter how many messages get queued up, our batch actor only makes two writes. Furthermore, if there are more than two Rotate commands in the same batch, earlier Text log lines can be discarded without being written at all! Together this can greatly improve the performance of working with external APIs.

Note that when extending BatchActor, it is up to the implementer to ensure that the BatchActors runBatch method has the same visible effect as if they had run a single run method on each message individually. Violating that assumption may lead to weird bugs, where the actor behaves differently depending on how the messages are batched (which is nondeterministic, and may depend on thread scheduling and other performance related details).

Debounced Logging using State Machines

The last common API we will look at is using StateMachineActor. We will define an actor that debounces writes to disk, ensuring they do not happen any more frequently than once every 50 milliseconds. This is a common pattern when working with an external API that you do not want to overload with large numbers of API calls.

sealed trait Msg
case class Flush() extends Msg
case class Text(value: String) extends Msg

class Logger(log: os.Path, debounceTime: java.time.Duration)
            (implicit ac: castor.Context) extends castor.StateMachineActor[Msg]{
  def initialState = Idle()
  case class Idle() extends State({
    case Text(value) =>
      ac.scheduleMsg(this, Flush(), debounceTime)
      Buffering(Vector(value))
  })
  case class Buffering(buffer: Vector[String]) extends State({
    case Text(value) => Buffering(buffer :+ value)
    case Flush() =>
      os.write.append(log, buffer.mkString(" ") + "\n", createFolders = true)
      Idle()
  })
}

implicit val ac = new castor.Context.Test()

val logPath = os.pwd / "out" / "scratch" / "log.txt"

val logger = new Logger(logPath, java.time.Duration.ofMillis(50))

In this example, we use StateMachineActor to define a Logger actor with two states Idle and Buffering.

This actor starts out with its initalState = Idle(). When it receives a Text message, it schedules a Flush message to be sent 50 milliseconds in the future, and transitions into the Buffering state. While in Buffering, any additional Text messages are simply accumulated onto the buffer, until the Flush is received again and all the buffered messages are flushed to disk. Each group of messages is written as a single line, separated by newlines (just so we can see the effect of the batching in the output). The output is as follows:

logger.send(Text("I am cow"))
logger.send(Text("hear me moo"))
Thread.sleep(100)
logger.send(Text("I weight twice as much as you"))
logger.send(Text("And I look good on the barbecue"))
Thread.sleep(100)
logger.send(Text("Yoghurt curds cream cheese and butter"))
logger.send(Text("Comes from liquids from my udder"))
logger.send(Text("I am cow, I am cow"))
logger.send(Text("Hear me moo, moooo"))

ac.waitForInactivity()

os.read.lines(logPath) ==> Seq(
  "I am cow hear me moo",
  "I weight twice as much as you And I look good on the barbecue",
  "Yoghurt curds cream cheese and butter Comes from liquids from my udder I am cow, I am cow Hear me moo, moooo",
)

You can see that when sending the text messages to the logger in three groups separated by 100 millisecond waits, the final log file ends up having three lines of logs each of which contains multiple messages buffered together.

In general, StateMachineActor is very useful in cases where there are multiple distinct states which an Actor can be in, as it forces you explicitly define the states, the members of each state, as well as the state transitions that occur when each state receives each message. When the number of distinct states grows, StateMachineActor can be significantly easier to use than SimpleActor.

While it is good practice to make your States immutable, StateMachineActor does not enforce it. Similarly, it is generally good practice to avoid defining "auxiliary" mutable state vars in the body of a StateMachineActor. The library does not enforce that either, but doing so somewhat defeats the purpose of using a StateMachineActor to model your actor state in the first place, in which case you might as well use SimpleActor.

Note that while multiple threads can send messages to Logger at once, and the Flush() message can also be sent at an arbitrary time in the future thanks to the ac.scheduleMsg call, the actor will only ever process one message at a time. This means you can be sure that it will transition through the two states Idle and Buffering in a straightforward manner, without worrying about multiple threads executing at once and messing up the simple state machine.

Debugging Actors

Debug Logging State Machines

When using StateMachineActor, all your actor's internal state should be in the single state variable. You can thus easily override def run to print the state before and after each message is received:

override def run(msg: Msg): Unit = {
  println(s"$state + $msg -> ")
  super.run(msg)
  println(state)
}

If your StateMachineActor is misbehaving, this should hopefully make it easier to trace what it is doing in response to each message, so you can figure out exactly why it is misbehaving:

logger.send(Text("I am cow"))
// Idle() + Text(I am cow) -> 
// Buffering(Vector(I am cow))
logger.send(Text("hear me moo"))
// Buffering(Vector(I am cow)) + Text(hear me moo) -> 
// Buffering(Vector(I am cow, hear me moo))
Thread.sleep(100)
// Buffering(Vector(I am cow, hear me moo)) + Debounced() -> 
// Idle()
logger.send(Text("I weight twice as much as you"))
// Idle() + Text(I weight twice as much as you) -> 
// Buffering(Vector(I weight twice as much as you))
logger.send(Text("And I look good on the barbecue"))
// Buffering(Vector(I weight twice as much as you)) + Text(And I look good on the barbecue) -> 
// Buffering(Vector(I weight twice as much as you, And I look good on the barbecue))
Thread.sleep(100)
// Buffering(Vector(I weight twice as much as you, And I look good on the barbecue)) + Debounced() -> 
// Idle()
logger.send(Text("Yoghurt curds cream cheese and butter"))
// Idle() + Text(Yoghurt curds cream cheese and butter) -> 
// Buffering(Vector(Yoghurt curds cream cheese and butter))
logger.send(Text("Comes from liquids from my udder"))
// Buffering(Vector(Yoghurt curds cream cheese and butter)) +
// Text(Comes from liquids from my udder) -> Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder))
logger.send(Text("I am cow, I am cow"))
// Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder)) + Text(I am cow, I am cow) -> 
// Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder, I am cow, I am cow))
logger.send(Text("Hear me moo, moooo"))
// Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder, I am cow, I am cow)) + Text(Hear me moo, moooo) -> 
// Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder, I am cow, I am cow, Hear me moo, moooo))

ac.waitForInactivity()
// Buffering(Vector(Yoghurt curds cream cheese and butter, Comes from liquids from my udder, I am cow, I am cow, Hear me moo, moooo)) + Debounced() ->
// Idle()

Logging every message received and processed by one or more Actors may get very verbose in a large system with lots going on; you can use a conditional if(...) in your override def run to specify exactly which state transitions on which actors you care about (e.g. only actors handling a certain user ID) to cut down on the noise:

override def run(msg: Msg): Unit = {
  if (???) println(s"$state + $msg -> ")
  super.run(msg)
  if (???) println(state)
}

Note that if you have multiple actors sending messages to each other, by default they run on a thread pool and so the println messages above may become interleaved and hard to read. To resolve that, you can try Running Actors Single Threaded.

Debugging using Context Logging

Apart from logging individual Actors, you can also insert logging into the castor.Context to log certain state transitions or actions. For example, you can log every time a message is run on an actor by overriding the reportRun callback:

implicit val ac = new castor.Context.Test(){
  override def reportRun(a: Actor[_], msg: Any, token: castor.Context.Token): Unit = {
    println(s"$a <- $msg")
    super.reportRun(a, msg, token)
  }
}

Running this on the two-actor pipeline example from earlier, it helps us visualize exactly what our actors are going:

castor.JvmActorsTest$Logger$5@4a903c98 <- I am cow
castor.JvmActorsTest$Logger$5@4a903c98 <- hear me moo
castor.JvmActorsTest$Logger$5@4a903c98 <- I weight twice as much as you
castor.JvmActorsTest$Writer$2@3bb87fa0 <- SSBhbSBjb3c=
castor.JvmActorsTest$Logger$5@4a903c98 <- And I look good on the barbecue
castor.JvmActorsTest$Logger$5@4a903c98 <- Yoghurt curds cream cheese and butter
castor.JvmActorsTest$Logger$5@4a903c98 <- Comes from liquids from my udder
castor.JvmActorsTest$Logger$5@4a903c98 <- I am cow, I am cow
castor.JvmActorsTest$Logger$5@4a903c98 <- Hear me moo, moooo
castor.JvmActorsTest$Writer$2@3bb87fa0 <- aGVhciBtZSBtb28=
castor.JvmActorsTest$Writer$2@3bb87fa0 <- SSB3ZWlnaHQgdHdpY2UgYXMgbXVjaCBhcyB5b3U=
castor.JvmActorsTest$Writer$2@3bb87fa0 <- QW5kIEkgbG9vayBnb29kIG9uIHRoZSBiYXJiZWN1ZQ==
castor.JvmActorsTest$Writer$2@3bb87fa0 <- WW9naHVydCBjdXJkcyBjcmVhbSBjaGVlc2UgYW5kIGJ1dHRlcg==
castor.JvmActorsTest$Writer$2@3bb87fa0 <- Q29tZXMgZnJvbSBsaXF1aWRzIGZyb20gbXkgdWRkZXI=
castor.JvmActorsTest$Writer$2@3bb87fa0 <- SSBhbSBjb3csIEkgYW0gY293
castor.JvmActorsTest$Writer$2@3bb87fa0 <- SGVhciBtZSBtb28sIG1vb29v

Running Actors Single Threaded

We can also replace the default scala.concurrent.ExecutionContext.global executor with a single-threaded executor, if we want our Actor pipeline to behave 100% deterministically:

implicit val ac = new castor.Context.Test(
  scala.concurrent.ExecutionContext.fromExecutor(
    java.util.concurrent.Executors.newSingleThreadExecutor()
  )
){
  override def reportRun(a: Actor[_], msg: Any, token: castor.Context.Token): Unit = {
    println(s"$a <- $msg")
    super.reportRun(a, msg, token)
  }
}

Any asynchronous Actor pipeline should be able to run no a newSingleThreadExecutor. While it would be slower than running on the default thread pool, it should make execution of your actors much more deterministic - only one actor will be running at a time - and make it easier to track down logical bugs without multithreaded parallelism getting in the way.

Changelog

0.3.0

  • Update sourcecode to 0.3.0
  • Drop support for Scala 2.11

0.1.7

  • Remove usage of scala.concurrent.ExecutionContext.global in favor of instantiating our own thread pool for castor.Context.Simple.global

0.1.1

  • Fix a NullPointerException when trying to use StateMachineActor with singleton objects

0.1.0

  • First release