Packet is a pure ruby library for writing network applications in Ruby. It follows Evented Model of network programming and implements almost all the features provided by EventMachine.

It also provides real easy to user UNIX workers for concurrent programming.

Its best to have some examples going:

Examples

A Simple Echo Server:

require “rubygems” require “packet”

class Foo

def receive_data p_data
  send_data(p_data)
end

def post_init
  puts "Client connected"
end

def connection_completed
  puts "Whoa man"
end

def unbind
  puts "Client Disconnected"
end

end

Packet::Reactor.run do |t_reactor|

t_reactor.start_server("localhost",11006,Foo)

end

Those new to network programming with events and callbacks, will note that, each time a new client connects an instance of class Foo is instantiated. When client writes some data to the socket, receive_data method is invoked.

Although Packet implements an API similar to EventMachine, but it differs slightly because of the fact that, for a packet app, there can be more than one reactor loop running and hence, we don’t use Packet.start_server(…).

A Simple Http Client

class WikiHandler

def receive_data p_data
  p p_data
end

def post_init
end

def unbind
end

def connection_completed
  send_data("GET / \r\n")
end

end

Packet::Reactor.run do |t_reactor|

t_reactor.connect("en.wikipedia.org",80,WikiHandler)

end

Using Workers

Packet enables you to write simple workers, which will run in 
different process and gives you nice 
evented handle for concurrent execution of various tasks. 

When, you are writing a scalable networking application 
using Event Model of network programming,
sometimes when processing of certain events take time, 
your event loop is stuck there. With green
threads, you don't really have a way of paralleling 
your request processing. Packet library, allows
you to write simple workers, for executing long 
running tasks. You can pass data and callbacks as an
argument.

When you are going to use workers in
your application, you need to define
constant WORKER_ROOT,
which is the directory location, where
your workers are located. All the workers defined in that directory
will be automatically, picked and forked in a
new process when your packet app starts. So, a typical
packet_app, that wants to use workers, will look like this:

packet_app_root
  |
  |__  lib
  |
  |___ worker
  |
  |___ config
  |
  |___ log

You would define WORKER_ROOT = PACKET_APP_ROOT/worker

All the workers must inherit class Packet::Worker, and hence a 
general skeleton of worker will look like:

  class FooWorker < Packet::Worker
    set_worker_name :foo_worker #=> This is necessary.
    def receive_data p_data
    end

    def connection_completed
    end

    def unbind
    end

    def post_init
    end
  end

All the forked workers are connected to master via
UNIX sockets, and hence messages passed to workers from master
will be available in receive_data method. Also,
when you are passing messages to workers, or worker is passing 
message to master ( in a nutshell, all the internal
communication between workers and master ) directly takes
place using ruby objects. All the passed ruby objects are
dumped and marshalled across unix sockets in a non blocking 
manner. BinParser class parses dumped binary objects and
makes sure, packets received at other end are complete.
Usually, you wouldn't need to worry about this little detail.

Packet provides various ways of interacting with
workers. Usually, when a worker is instantiated, a proxy for
that worker will also be instantiated at master
process. Packet automatically provides a worker proxy(See meta_pimp.rb)
for you, but  if you need to multiplex/demultiplex
requests based on certain criteria, you may as well define your 
own worker proxies. Code, would like something like this:

  class FooWorker < Packet::Worker
    set_worker_proxy :foo_handler
  end

When you define, :foo_handler as a proxy for
this worker, packet is gonna search for FooHandler class and 
instantiate it when the worker gets started. All 
the worker proxies must inherit from Packet::Pimp. 
Have a look at, Packet::MetaPimp,
which acts as a meta pimp for all the workers,
which don't have a explicit worker proxy defined.

A complete Case :

 Just for kicks, lets write a sample server,
 which evals whatever clients send to it. But, assuming this 'eval' of 
 client data can be potentially time/cpu
 consuming ( not to mention dangerous too ), we are gonna ask our eval_worker, to
 perform eval and return the result to master process, which in 
 turn returns the result to happy client.

 # APP_ROOT/bin/eval_server.rb
 EVAL_APP_ROOT = File.expand_path(File.join(File.dirname(__FILE__) + "/.."))
 ["bin","worker","lib"].each { |x| $LOAD_PATH.unshift(EVAL_APP_ROOT + "/#{x}")}
 WORKER_ROOT = EVAL_APP_ROOT + "/worker"

 require "packet"
 class EvalServer
   def receive_data p_data
     ask_worker(:eval_worker,:data => p_data, :type => :request)
   end

   # will be called, when any worker sends data back to master process
   # it should be noted that, you may have several instances of eval_server in
   # your master, for each connected client, but worker_receive will be always
   # be invoked for the instance, which originally made the request.
   # If you need fine control, over this behaviour, you can implement a worker proxy 
   # on the lines of meta_pimp class. This API will change in future perhaps, as i
   # expect, better ideas to come.  
   def worker_receive p_data
     send_data "#{p_data[:data]}\n"
   end

   def show_result p_data
     send_data("#{p_data[:response]}\n")
   end

   def connection_completed
   end

   def post_init
   end

   def wow
     puts "Wow"
   end
 end

 Packet::Reactor.run do |t_reactor|
   t_reactor.start_server("localhost", 11006,EvalServer) do |instance|
     instance.wow
   end
 end

# APP_ROOT/worker/eval_worker.rb
  class EvalWorker < Packet::Worker
  set_worker_name :eval_worker
  def worker_init
    p "Starting no proxy worker"
  end

  def receive_data data_obj
    eval_data = eval(data_obj[:data])
    data_obj[:data] = eval_data
    data_obj[:type] = :response
    send_data(data_obj)
  end
end

Disable auto loading of certain workers:

Sometimes, you would need to start a
worker at runtime and don't want this pre-forking mechanism.
Packet, allows this. You just need to define
"set_no_auto_load true" in your worker class and worker
will not be automatically forked. Although name is a bit misleading perhaps.

Now, at runtime, you can call start_worker(:foo_worker, options)
to start a worker as usual. It should
be noted that, forking a worker, which is already
forked can be disastrous, since worker names are being
used as unique keys that represent a worker.Test

Performance:

Although written in pure ruby, packet performs
reasonably well. Mongrel, running on top of Packet is a tad
slower than Mongrel running on top of EventMachine. More benchmarks coming soon.

SVN repo:

Code for packet is on google code, svn repo is:
http://packet.googlecode.com/svn/trunk/

Credits

Francis for awesome EventMachine lib, which has constantly acted as an inspiration.
Ezra, for being a early user and porting mongrel to run on top of packet