Thursday, June 04, 2009

Liveblogging of Scala Actors Talk from JavaOne

Today I was attending a talk on Scala Actor by Phillipp Haller and Frank Sommers. Tyler asked me to live blog it. I thought that was a good idea, and that Twitter would be the way to do it. This did not work out, as Twitter shut me down for too many tweets in a given hour. So now I am posting the full live blog, both the tweets that went through and the ones that Twitter rejected. However, pulling out the text of all of the good tweets is not totally straightforward. I figured the easiest was to do this was to use some Scala with the Twitter API via the REPL:


scala> import java.net._
import java.net._

scala> import scala.xml._
import scala.xml._

scala> val url = new URL("http://twitter.com/statuses/user_timeline.xml?since_id=2035817713&max_id=2036208882&screen_name=michaelg&url: java.net.URL = http://twitter.com/statuses/user_timeline.xml?since_id=2035817713&max_id=2036208882&screen_name=michaelg&count=75

scala> val conn = url.openConnection conn: java.net.URLConnection = sun.net.www.protocol.http.HttpURLConnection:http://twitter.com/statuses/user_timeline.xml?since_id=2035817713&max_id=2036208882&screen_name=michaelg&count=75

scala> val tweets = XML.load(conn.getInputStream) tweets: scala.xml.Elem =


scala> (tweets\\"text").foreach{ node => println(node.text) }


The result of the last call produces the following:

This allows for many more actors than worker threads #javaone
Work stealing makes this even more efficient. Let worker threads steal work from other worker threads #javaone
So decouple from threads using thread pools from java.util.concurrent. #javaone
A naive impl of thread-per-actor has overhead from thread creation, context-switching and memory use. #javaone
How event-based actors decouples threads and actors. How the execution enviro is setup. How the DSL works #javaone
Now time to go under the hood of Scala Actors #javaone
Only use receive when you must return a value from an Actor #javaone
Use react whenever possible and only use receive when necessary #javaone
When to use receive vs. react? #javaone
Replace while(true) { ... } with loop { ... } #javaone
Go from 5,000 actors/JVM to 1,000,000/JVM using event based Actors #javaone
Replace receive with react, receiveWithin with reactWithin, and while(condition) with loopWhile (condition) #javaone
To scale to a large number of Actors, go for event based Actors #javaone
The sender of a message is still sent across the network #javaone
Everything else works exactly the same as it would for local actors #javaone
Use the select function to get a proxy to the remote Actor #javaone
To make an actor remote is very easy. Call the alive function and the register function #javaone
So far everything has been within a single VM, but Scala Actors scales using remote actors #javaone
To specify timeouts, use receiveWithin. This takes a timeout and sends a TIMEOUT message #javaone
All of this makes it easy to keep track of subscribers and send them messages, like new chat messages #javaone
You can also wait for futures using receive function #javaone
When you try to access the future, it will then block if the reply has not been received #javaone
For async with assurance that the message was received using double-bang !! Returns a future representing the reply #javaone
Synchronous messages are allowed too using !? This blocks until reply received #javaone
Bang is asyc, returns immediately and passes a reference to the sender #javaone
To subscribe a user to the chat room, just send a message using the "bang" operator (method): ! #javaone
There is another API shortcut for creating an actor inline using the "actor" function #javaone
Creating a subscription is just a matter of sending message stating as much. Lets recipient capture the state #javaone
Common pattern of "child actors", benefits from asych communication of actors #javaone
Make each user an Actor as well to maintain state, such as who sent a message to the chat room #javaone
Now going back to the chat application to demonstrate Actor subscriptions #javaone
Patterns are tried in order, and the first match wins. No fall-through. Looks like a factory method call. #javaone
This is all based on Scala's pattern matching, think Java's switch statement but on steroids #javaone
If no message in mailbox, the actor will suspend until there is a message. Syntax for matching messages is super simple #javaone
Use the receive function to process messages within the act method of your Actor. Receive gets a message from the Actor's mailbox #javaone
Defining messages is made easy by using Scala's case classes. These are pure data classes typically. #javaone
Creating an actor is very similar to creating a thread in Java. #javaone
Creating an Actor is easy, just do 3 things. Extend Actor. Implement act method. Start the actor. #javaone
Frank showing a chat application architecture that uses Actors "the quintessential Actor system" #javaone
Actor implemented entirely as a DSL, but part of Scala std. library. Used in big systems such as Twitter #javaone
Actors can be local or remote, no change to programming model. #javaone
Actors are event based in Scala, not tied to Java threads. So much more lightweight. A single JVM can have millions of actors #javaone
Actors use several Scala language features: pattern matching, closures, traits/multiple inheritance, and Java interop. #javaone
Actors in Scala are probably the closest implementation of Erlang model on the JVM #javaone
OTP: library of Erlang modules for building reliable distributed systems #javaone
Erlang has support for Actors built-in. Erlang VM is process-oriented, makes Actor model very scalable. #javaone
Best known example of Actors though is from Erlang, "a pure message passing language" #javaone
Actors are an established pattern. Research dates to 70s at MIT in Smalltalk and Simula #javaone
Mutable messages are also possible. Actors can also be local or remote -- same programing model #javaone
Actors share nothing, private state. Thus no synchronization. Actors send messages, but the message are immutable #javaone
Actors are active objects and has a mailbox. Actors consume messages, perform actions but the code is sequential within an actor #javaone
There is an existing concurrency model that fits with the described solutions: Actors. #javaone
The solution: DSL + sequential programs. Shared nothing. Message passing. Asynchronous communication #javaone
Why learn about Scala actors? Concurrency presents opportunity and problems for developers, especially large teams of developers #javaone
Frank and Phillip are writing a book on Scala actors #javaone
Phillip Haller and Frank Sommers talking about Scala actors now #javaone

So that's the good tweets. Here are the ones that did not go through...


The actor method takes a closure creates an Actor
The react method saves pattern matching as a closure to be executed later
Uses exception for control flow, any code after react will not execute
Actors usually execute on different threads, but you can control what threads
This is useful for threadLocals and for Swing
A demo of SwingActor
Example of using link and trapExit, but not much detail of this shown
Scala Actors provide a simple programming model to write concurrent code on the JVM
Many projects already use actors: Twitter, Lift, Scala OTP (@jboner), partest
Future of Scala Actors in 2.8
Pluggable schedulers -- create actors with daemon-like behavior for example
Integrating exceptions and Erlang-style error handling. Get the best of both Java and Erlang error handling
Support for continuations. Currently requires optional compiler-plugin. Allows for more flexible control structures
Actor migrations (from machine to machine) could also be done with continuations
More static checking of messages, in particular an @immutable annotation
Actor isolution -- no accidentally sharing mutable state
This will allow for better passing mutable messages for performance benefits
No problem passing large methods, as they are passed by reference

1 comment:

Tyler Weir said...

>Pluggable schedulers -- create actors with daemon-like behavior for example

I wonder if this will help out with dpp's issues with Actors.