Oni

Oni is a Ruby framework that aims to make it easier to write concurrent daemons using a common code structure. Oni itself does not actually daemonize your code, manage PID files, resources, etc. Instead you should use Oni in combination with other Gems such as daemon-kit.

Oni was built to standardize the structure amongst the different daemon-kit projects used at Olery. As time progressed new structures were used for new daemons and the old ones were often left as is.

Another problem we faced was concurrency. Most daemons were built in a single threaded, single processed manner without an easy place to hook some kind of concurrency model in.

Oni takes care of these problems by providing the following:

  • A concurrency model in the form of separate worker threads (5 by default).
  • Clear separation of logic into 3 distinctive parts.
  • A common structure for your daemon projects.

Oni assumes developers are somewhat familiar with threading and the potential issues that may arise, it also assumes that your code doesn't leak large amounts of memory over time. Currently there are no plans to include some kind of internal resource management system in Oni, this may change in the future.

Design

To understand the design of Oni we'll first look at the typical work flow of a daemon:

  1. A job gets put in a queue (Amazon SQS).
  2. Daemon polls queue, takes message.
  3. Optional message format validation (this was rarely enforced since we had control over the input and assumed it to be correct).
  4. Work gets offloaded to some extra class, in older designs of our daemons it would happen in the daemon layer directly.
  5. Optionally the input data would be modified and re-queued into a separate queue. For example, we often pass along data through multiple queues from the start until the very end (e.g. batch IDs).

Oni tries to make these kind of workflows by breaking them up into 3 different layers:

  1. A daemon layer tasked with receiving and scheduling work.
  2. A "mapper" tasked with transforming (and validating) input/output for/from the worker. It sits between the daemon and the worker.
  3. A worker that actually performs a task (asynchronously)

Each layer would only do the specific thing it should be doing and would offload other work to the next step in the process. The 3 parts are described in detail below.

The Daemon

The daemon layer spawns a number of threads that will each receive and perform work separately. Typically these workers are long running tasks that poll some kind of message queue for jobs to process.

In initial iterations Oni used a main job dispatcher (running in the main thread) and a separate thread pool for the workers. This proved problematic with message queue setups as it would result in the main thread pulling in all available jobs and then internally queing them again given there weren't enough workers available. This would mean that if the process would crash the messages were lost. As a result of this each worker is started in it's own separate thread.

Comparing Oni with other framework structures one could see the daemon layer as a controller (in MVC frameworks), it merely dispatches work to the mapper and worker instead of doing everything itself.

The Mapper

The mapper is tasked with two things:

  1. Take the input from the daemon, validate it and transform it into a structure that the worker can understand.
  2. Take the resulting output, optionally modify it and pass it back to the daemon.

The input transformation is put in place to ensure that workers only get data that they actually need instead of just receiving the raw message (which may include all kinds of meta data completely useless to a worker). It also ensures that the input is actually correct before ever passing it to a worker.

A typical thing at Olery is that a job gets scheduled and has to pass through multiple steps (= daemons) before being completed. Some of these daemons would add extra meta-data to the message (e.g. batch IDs, timings, etc) but these aren't strictly required to perform an actual job, thus there's no need to pass it around several layers deep into your codebase.

In the above case the mapper would take care of validating/scrubbing the input and adding extra meta-data to the output.

The Worker

The worker would perform the actual work and return some kind of output to the daemon. Oni assumes that workers behave reasonably well, currently there's no mechanism in place to deal with memory leaks and the likes. Oni also assumes that developers are somewhat capable of dealing with asynchronous code since all work is performed asynchronously outside of the daemon layer.

Requirements

  • Ruby 1.9.3 or newer, preferably an implementation without a GIL such as Rubinius or Jruby.
  • Basic understanding of threading/concurrent programming

Installation & Basic Usage

Install the Gem:

gem install oni

Basic usage of Oni is as following:

require 'oni'

# This example defines 3 classes: MyDaemon, MyMapper and MyWorker.
# Combined they form the basic structure of an Oni Daemon.

class MyDaemon < Oni::Daemon
  set :mapper, MyMapper
  set :worker, MyWorker

  # Here you'd receive your message, e.g. from a queue. We'll use static
  # data as an example.
  def receive
    yield({:number => 10})
  end

  # This would get executed upon completion of a job.
  def complete(message, result)
    puts result
  end
end

class MyMapper < Oni::Mapper
  # Map the input given by MyDaemon#receive into the right arguments
  # for MyWorker#initialize.
  #
  # NOTE: the return value should be an Array.
  # Oni calls #to_a on the output (more specifically we use a splat) so your
  # return value should be an Array. If you do not return an Array you risk
  # that the object you return will wrongfully be converted into an Array.
  # This is painful when it comes to Hashes, which, after the #to_a call
  # will be converted into arrays of values in stead of your intended hash.
  def map_input(input)
    return [input[:number]]
  end

  def map_output(output)
    return {:number => output, :completed => Time.now}
  end
end

class MyWorker < Oni::Worker
  def initialize(number)
    @number = number
  end

  def process
    return @number * 2
  end
end

License

The source code of this repository and Oni itself are licensed under the MIT license unless specified otherwise. A copy of this license can be found in the file "LICENSE" in the root directory of this repository.