Nanite : A self-assembling fabric of Ruby daemons

Google Group: groups.google.com/group/nanite irc.freenode.net: #nanite

Intro

Nanite is a new way of thinking about building cloud ready web applications. Having a scalable message queueing back-end with all the discovery and dynamic load based dispatch that Nanite has is a very scalable way to construct web application back-ends.

A Nanite system has two types of components. There are nanite agents, these are the daemons where you implement your code functionality as actors. And then there are mappers.

Mappers are the control nodes of the system. There can be any number of mappers, these typically run inside of your merb or rails app running on the thin webserver (eventmachine is needed) but you can also run command line mappers from the shell.

Each Nanite agent sends a ping to the mapper exchange every @ping_time seconds. All of the mappers are subscribed to this exchange and they all get a copy of the ping with your status update. If the mappers do not get a ping from a certain agent within a timeout @ping_time the mappers will remove that agent from any dispatch. When the agent comes back online or gets less busy it will re-advertise itself to the mapper exchange therefore adding itself back to the dispatch. This makes for a very nice self-healing cluster of worker processes and a cluster of front-end mapper processes.

In your Nanites you can have any number of actor classes. These actors are like controllers in Rails or Merb and this is where you implement your own custom functionality. An actor looks like:

class Foo
  include Nanite::Actor
  expose :bar
  def bar(payload)
    "got payload: #{payload}"
  end
end
Nanite::Dispatcher.register(Foo.new)

The methods that you ‘expose’ on the actors are advertised to the mappers like:

Foo#bar   => /foo/bar
Mock#list => /mock/list

Every agent advertises its status every time it pings the mapper cluster. The default status that is advertised is the load average as a float. This is used for the default request dispatching based on least loaded server.

You can change what is advertised as status to anything you want that is comparable(<=>) by doing something like this in your agent’s init.rb file:

status_proc = lambda { MyApp.some_statistic_indicating_load }

This proc will be recalled every @ping_time and sent to the mappers.

This is the ‘fitness function’ for selecting the least loaded Nanite. You can dream up any scheme of populating this function so the mappers can select the right Nanites based on their status.

A quick note about security:

Nanite security is based upon RabbitMQ vhosts so anything attached to one vhost can talk to anything else on the same vhost. So you generally want one vhost per app space.

Installation

Nanite has a lot of moving parts. Follow the directions below and we’ll get you up and running in no-time.

Install Erlang (OS X)

See the Erlang website for the latest info : www.erlang.org/download.html

In your chosen source dir which we’ll refer to as <SRC>:

cd <SRC>
wget http://www.erlang.org/download/otp_src_R12B-5.tar.gz
tar -zxvf otp_src_R12B-5.tar.gz
cd otp_src_R12B-5
./configure --enable-hipe --enable-darwin-universal
make
sudo make install

Install Erlang (Linux .deb)

sudo apt-get install erlang-nox

Install Nanite Ruby Gem

Installing the gem gives us access to the various Nanite commands in the default binary path.

cd <SRC>
git clone git://github.com/ezmobius/nanite.git
cd nanite
rake gem
sudo gem install pkg/nanite-<VERSION>.gem

Install EventMachine Ruby Gem

sudo gem install eventmachine

Install AMQP Ruby Gem

# install version >= 0.6.0 from RubyForge
sudo gem install amqp

# or install from source
cd <SRC>
git clone git://github.com/tmm1/amqp.git
cd amqp && rake gem && sudo gem install amqp-<VERSION>.gem

Install RabbitMQ from source tarball (OS X and generic Linux)

In short, you’ll need a working erlang installation, python, and simplejson. If your python (“python -V”) is 2.6, simplejson is included in your distribution. Otherwise, install simplejson:

easy_install simplejson

If you don’t have easy_install, install setuptools from pypi.python.org/pypi/setuptools:

# Get the latest .egg file for your version of python ("python -V"), then run it:
wget http://pypi.python.org/packages/2.5/s/setuptools/setuptools-0.6c9-py2.5.egg
sh setuptools-0.6c9-py2.5.egg

and install simplejson, which is needed for the “make” step below.

These instructions assume the latest RabbitMQ release 1.5.3:

# Download somewhere
cd /root
wget http://www.rabbitmq.com/releases/rabbitmq-server/v1.5.3/rabbitmq-server-1.5.3.tar.gz

# Go to your erlang lib directory, usually /usr/lib/erlang/lib or:
cd /usr/local/lib/erlang/lib

tar -zxf ~/rabbitmq-server-1.5.3.tar.gz
cd rabbitmq-server-1.5.3
make

# There is no "make install" phase.

Be sure to add the /usr/local/lib/erlang/lib/rabbitmq-server-1.5.3/scripts to your $PATH.

The following websites may also be useful:

Install RabbitMQ (Linux .deb)

wget http://www.rabbitmq.com/releases/rabbitmq-server/v1.5.3/rabbitmq-server_1.5.3-1_all.deb
sudo apt-get install logrotate
sudo dpkg -i rabbitmq-server_1.5.3-1_all.deb

Test your installation

Test your Erlang install

Start an Erlang shell

erl

Enter the following commands in the Erlang shell (When installed correctly each should print a great deal of erlang config info):

rabbit:module_info().

Exit the Erlang shell with the following command (or Ctrl-c):

q().

Start RabbitMQ

All directories will be automatically set up on first run.

To run RabbitMQ in the foreground:

sudo rabbitmq-server

To run RabbitMQ in the background:

sudo rabbitmq-server -detached

To check status:

rabbitmqctl status

To stop server:

rabbitmqctl stop

You can learn more about RabbitMQ admin here: www.rabbitmq.com/admin-guide.html

Test Ruby/EventMachine/AMQP end-to-end with a running RabbitMQ instance

You can test that all the moving parts are working end-to-end by running one of the AMQP example programs. There are a number of example tests provided with AMQP but the one we will try simply queues and prints two messages immediately, and another one five seconds later.

# this test can only be run if you made a local clone
# of the amqp repository during the amqp installation above
# (it does not matter however if the gem was installed from src or rubyforge)
cd <SRC>/amqp
ruby examples/mq/simple-get.rb

Test Nanite (finally)

First we’ll do a little setup and run a script which adds an agent account (nanite), a mapper account (mapper) and a ‘/nanite’ vhost to RabbitMQ. RabbitMQ broker must of course be running before you run this script.

cd nanite
./examples/rabbitconf.rb

Now lets run a few agents. Each of these is a long running process and needs to run in its own shell. Each agent also needs its own unique identity, or token to differentiate it from the other running agents. You can try opening a couple of them now if you like. The agents will generally provide no output when running. You can always get more help about the nanite command with ‘nanite –help’.

First shell

cd examples/simpleagent
nanite-agent --token fred

Second shell

cd examples/simpleagent
nanite-agent --token bob

Now run a mapper. Mappers can be run from within your Merb or Rails app, from an interactive irb shell, or from the command line. For this example we’ll run it from the command line so open a third shell window and run the following:

cd examples
./cli.rb

Which should soon return something like the following.

{"bob"=>"hello nanite"}                                 # where the '--token bob' parameter was passed
{"55a7f300c454203eacc218b6fbd2edc6"=>"hello nanite"}    # where no '--token' was passed.  auto-generated identity.

Now if you want to make this interesting, you can issue a Ctrl-c in one of the agent’s windows to kill it. And then run cli.rb again. You should see that you still get a result back since the mapper is finding the remaining agent to do its work.

If you want to try requesting work to be done by an agent from an interactive command shell try the following. This assumes that you have the agents running as indicated in the example above (‘>>’ is the nanite shell prompt).

cd nanite
./bin/nanite-mapper -i -u mapper -p testing -v /nanite
Starting mapper console
>> request('/simple/echo') {|res| p res }

By default this will dispatch to the agent with the lowest reported load average.

There are a few other selectors as well:

# run this request on *all* agents that expose the /foo/bar Foo#bar actor
>> request('/foo/bar', 'hi', :selector => :all) {|res| p res }

# run this request on one random agent that expose the /whatever/hello Whatever#hello actor
>> request('/whatever/hello', 42, :selector => :random) {|res| p res }

You can create your own selectors based on arbitrary stuff you put in status from your agents see cluster.rb for examples of how least_loaded, all and random are implemented.

You can run as many mappers as you want, they will all be hot masters.

The calls are asynchronous. This means the block you pass to Nanite::Agent#request is not run until the response from the agent(s) have returned. So keep that in mind. Should you need to poll from an ajax web app for results you should have your block stuff the results in the database for any web front end to pick up with the next poll.

Another option to test your agents is to use nanite-admin

$ nanite-admin 
starting nanite-admin
>> Thin web server (v1.0.1 codename ?)
>> Maximum connections set to 1024
>> Listening on 0.0.0.0:4000, CTRL+C to stop

and bring up the Nanite Control Tower at localhost:4000 .

Have fun!

Web Framework Integration

The integration of Nanite into web frameworks depends on the web server running our application and, more importantly, if it uses EventMachine or not.

Thin is EventMachine-based, so we only need to make sure that the EventMachine reactor is already ‘heated’ up

Thread.new do
  until EM.reactor_running?
    sleep 1
  end
  Nanite.start_mapper(:host => 'localhost', :user => 'mapper', :pass => 'testing', :vhost => '/nanite', :log_level => 'info')
end

Mongrel on the other hand does not use EventMachine and therefore requires to wrap the start of our mapper

Thread.new do
  EM.run do 
    Nanite.start_mapper(:host => 'localhost', :user => 'mapper', :pass => 'testing', :vhost => '/nanite', :log_level => 'info')
  end
end

Where to put the mapper initialization code depends on the framework and our preference. For Rails the canonical place to start our mapper is within nanite.rb (or any other filename you prefer) in config/initalizers. In Merb we can use init.rb in config.

Troubleshooting

** IMPORTANT ** If you are using Apple’s built in Ruby that comes with Leopard (10.5.x) or Tiger (10.4.x) then your READLINE lib is hosed and the interactive shell will not work. As soon as you drop into a shell Apple’s fakey READLINE will halt the event machine loop so you won’t see any nanites registering. I don’t have a good workaround except to tell you not to use Apple’s Ruby, build your own or use ports.

** What to do if rabbitconf.rb dies with: badrpc,nodedown and nothing you do seems to matter **

If rabbitconf.rb dies saying “badrpc,nodedown” it means that for some reason, the rabbitmqctl program rabbitconf.rb is using to setup the agent accounts for nanite is not able to connect to your RabbitMQ server. Assuming RabbitMQ is running and is known to work (try the examples that come with the amqp library), then there’s a chance something is going wrong with using short node names in your Erlang install.

The easiest way to verify it is by starting two separate Erlang shells like this (note that “odin” is my hostname):

$ erl -sname fred
(fred@odin)1>

$ erl -sname bob
(bob@odin)1>

And then trying to ‘ping’ one from the other. In the ‘fred’ node you can do that like this:

(fred@odin)1> net_adm:ping(bob@odin).
pang

Note : If your hostname has a hyphen in it (e.g. macbook-pro) you should include the user@hostname in quotes like:

net_adm:ping('bob@macbook-pro').

If you see ‘pang’ (which is apparently Swedish for something like “crap! it’s broken.”) then short name distribution isn’t working for you, and you need to fall back to using a full name. If you see ‘pong’, then that’s not actually your problem.

First, verify that your system’s hostname is set to a fully qualified domain name. On OS X it can end in ‘.local’:

hostname
odin.local

Then test that this will work by starting a node with the full name of ‘fred@<your hostname>’:

erl -name [email protected]
([email protected])1>

then bob@<yourhostname>, and then finally try to ping fred:

erl -name [email protected]
([email protected])1> net_adm:ping([email protected]).
pong

In my case, it looks like that worked. Now… on to getting rabbitconf.rb to run! To do that, you need to edit the ‘rabbitmq-server’ and ‘rabbitmqctl’ scripts in your RabbitMQ distribution and edit the -sname arguments to use -name and a full name.