No commit activity in last 3 years
No release in over 3 years
JobReactor is a library for creating, scheduling and processing background jobs. It is asynchronous client-server distributed system based on EventMachine.
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
 Dependencies

Runtime

 Project Readme

JobReactor

JobReactor is a library for creating, scheduling and processing background jobs. It is asynchronous client-server distributed system based on EventMachine.

JobReactor is the best solution for I/O intensive web application powered by evented web servers such Thin. While all requests are processed in one thread you should avoid blocking reactor loop. So you should (and must) delegate intensive calculations to another process.

Use JobReactor to avoid blocking calculations.

JobReactor keeps you in the evented paradigm by allowing to register callback for the tasks which will be triggered when work is done. See simple example of using with AsyncSinatra here.

To use JobReactor with Sinatra or Ruby on Rails you should start distributor in initializer using JR.run method (it launches EventMachine in separate thread). Then add rake task(s) which will run the node(s).

If you use server based on EventMachine (Thin) use JR.wait_em_and_run method which will initialize JobReactor when EventMachine started.

So, read the 'features' section and try JobReactor. You can do a lot with it.

Note

JobReactor is based on EventMachine. Jobs are launched in EM reactor loop in one thread. There are advantages and disadvantages. The main benefit is fast scheduling, saving and loading. The weak point is the processing of heavy background jobs when each job takes minutes and hours. They will block the reactor and break normal processing.

If you can't divide 'THE BIG JOB' into 'small pieces' you shouldn't use JobReactor. See alternatives such DelayedJob or Resque.

JobReactor is the right solution if you have thousands, millions, and, we hope, billions relatively small jobs.

Quick start

gem install job_reactor

You should install Redis if you want to persist your jobs.

$ sudo apt-get install redis-server

In your main application: application.rb

require 'job_reactor'

JR.run do
  JR.start_distributor('localhost', 5000)  #see lib/job_reactor/job_reactor.rb
end

sleep(1) until(JR.ready?)

# The application
loop do
  sleep(3) #Your application is working
  JR.enqueue 'my_job', {arg1: 'Hello'}
end

Define the 'my_job' in separate directory (files with job's definitions must be in separate directory): reactor_jobs/my_jobs.rb

include JobReactor

job 'my_job' do |args|
  puts args[:arg1]
end

And the last file - 'the worker code': worker.rb

require 'job_reactor'

JR.config[:job_directory] = 'reactor_jobs' #this default config, so you can omit this line

JR.run! do
  JR.start_node({
  :storage => 'memory_storage',
  :name => 'worker_1',
  :server => ['localhost', 5001],
  :distributors => [['localhost', 5000]]
  })                                         #see lib/job_reactor/job_reactor.rb
end

Run 'application.rb' in one terminal window and 'worker.rb' in another. Node connects to distributor, receives the job and works. Cool! But it was the simplest example. See 'examples' directory.

Features

  1. Client-server architecture

You can run as many distributors and working nodes as you need. You are free to choose the strategy. If you have many background tasks from each part of your application you can use, for example, 3 distributors (one in each process) and 10 working nodes. If you don't have many jobs you can leave only one node which will be connected to 3 distributors. 2. High scalability

Nodes and distributors are connected via TCP. So, you can run them on any machine you can connect to. Nodes may use different storage or the same one. You can store vitally important jobs in database and simple insignificant jobs in memory. And more: You can run node and distributor inside one EMreactor, so your nodes may create jobs for others nodes and communicate with each other. 3. Full job control

You can add 'callback' and 'errbacks' to the job which will be called on the node. You also can add 'success feedback' and 'error feedback' which will be called in your main application. When job is done on remote node, your application will receive the result inside corespondent 'feedback'. If error occur in the job you can see it in 'errbacks' and the in 'error feedback' and do what you want. 4. Reflection and modifying

Inside the job you can get information about when it starts, when it fails, which node execute job and etc. You also can add some arguments to the job on-the-fly which will be used in the subsequent callbacks and errbacks. These arguments then can be sent back to the distributor. 5. Reliability

You can run additional nodes and stop any nodes on-the-fly. Distributor is smart enough to send jobs to another node if someone is stopped or crashed. If no nodes are connected to distributor it will keep jobs in memory and send them when nodes start. If node is stopped or crashed it will retry stored jobs after start. 6. EventMachine available

Remember, your jobs will be run inside EventMachine reactor! You can easily use the power of async nature of EventMachine. Use asynchronous em-http-request, em-websocket, and etc. 7. Thread safe

EventMachine reactor loop runs in one thread. So the code in jobs executed in the given node is absolutely threadsafe. The only exception is 'defer' job, when you tell the node to run job in EM.defer block (so job will be executed in separate thread). 8. Deferred and periodic jobs

You can use deferred jobs which will run 'after' some time or 'run_at' given time. You can create periodic jobs which will run every given time period and cancel them on condition. 9. No polling

There is no storage polling. Absolutely. When node receives job (no matter instant, periodic or deferred) there will be EventMachine timer created which will start job at the right time. 10. Job retrying

If job fails it will be retried. You can choose global retrying strategy or manage separate jobs. 11. Predefined nodes

You can specify node for jobs, so they will be executed in that node environment. And you can specify which node is forbidden for the job. If no nodes are specified distributor will try to send the job to the first free node. 12. Node based priorities

There are no priorities like in Delayed::Job or Stalker. But there are flexible node-based priorities. You can specify the node which should execute the job and the node is forbidden for given job. You can reserve several nodes for high priority jobs.

How it works

  1. You run JobReactor::Distributor in your application initializer

JR.run do
  JR.start_distributor('localhost', 5000)
end

This code runs EventMachine reactor loop in the new thread and call the block given. JR.start_distributor starts EventMachine TCP server on given host and port. And now JobReactor is ready to work.

  1. You run JobReactor::Node in the different process or different machine

JR.run! do
  JR.start_node({
    storage: 'redis_storage',
    name: 'redis_node1',
    server: ['localhost', 5001],
    distributors: [['localhost', 5000]] 
})
end

This code runs EventMachine reactor loop (in the main thread: this is the difference between run and run!). And start the Node inside the reactor. When node starts it:

  • parses the 'reactor jobs' files (recursively parse all files specified in JR.config[:job_directory] directory, default is 'reactor_jobs' directory) and create hash of jobs callbacks and errbacs (see [JobReator jobs]);
  • tries to 'retry' the job (if you use 'redis_storage' and JR.config[:retry_jobs_at_start] is true)
  • starts it's own TCP server;
  • connects to Distributor server and sends the information about needed to establish the connection; When distributor receives the credentials it connects to Node server. And now there is a full duplex-connection between Distributor and Node.
  1. You enqueue the job in your application

JR.enqueue('my_job',{arg1: 1, arg2: 2}, {after: 20}, success, error)

The first argument is the name of the job, the second is the arguments hash for the job. The third is the options hash. If you don't specify any option job will be instant job and will be sent to any free node. You can use the following options:

  • defer: true or false - node will run the job in 'EM.defer' block. Be careful, the default thread pool size is 20 for EM. You can increase it by setting EM.threadpool_size = 'your value', but it is not recommended;
  • after: seconds - node will try run the job after seconds seconds;
  • run_at: time - node will try run the job at given time;
  • period: seconds - node will run job periodically, each seconds seconds; You can add node: 'node_name' and not_node: 'node_name' to the options. This specify the node on which the job should or shouldn't be run. For example:
JR.enqueue('my_job', {arg1: 1}, {period: 100, node: 'my_favourite_node', not_node: 'do_not_use_this_node'})

The rule to use specified node is not strict if JR.config[:always_use_specified_node] is false (default). This means that distributor will try to send the job to the given node at first. But if the node is locked (maybe you have just sent another job to it and it is very busy) distributor will look for other node.

The last two arguments are optional. The first is 'success feedback' and the last is 'error feedback'. We use term 'feedback' to distinguish from 'callbacks' and 'errbacks'. 'feedback' is executed on the main application side while 'callbacks' on the node side. 'feedbacks' are the procs which will be called when node sent message that job is completed (successfully or not). The argunments for the 'feedback' are the arguments of the initial job plus all added on the node side.

Example:

#in your 'job_file'
job 'my_job' do |args|
  #do smth
  args.merge!(result: 'Yay!')
end

#in your application
#success feedback
success = proc {|args| puts args}
#enqueue job
JR.enqueue('my_job', {arg1: 1}, {}, success)

The 'success' proc args will be {arg1: 1, result: 'Yay!'}. The same story is with 'error feedback'. Note, that error feedback will be launched after all attempts failed on the node side. See config: JR.config[:max_attempt] = 10 and JR.config[:retry_multiplier]

  1. You disconnect node (stop it manually or node fails itself)

  • distributor will send jobs to any other nodes if present
  • distributor will store in memory enqueued jobs if there is no connected node (or specified node)
  • when node starts again, then distributor will send jobs to the node
  1. You stop the main application.

  • Nodes will continue to work, but you won't be able to receive the results from node when you start the application again because all feedbacks are stored in memory.

Callbacks and feedbacks

'callbacks', 'errbacks', 'success feedback', and 'error feedback' helps you divide the job into small relatively independent parts.

To define 'job' you use JobReactor.job method (see 'Quick start' section). The only arguments are 'job_name' and the block which is the job itself.

You can define any number of callbacks and errbacks for the given job. Just use JobReactor.job_callback and JobRector.job_errback methods. The are three arguments for calbacks and errbacks. The name of the job, the name of callback/errback (optional) and the block.

include JobReactor

job 'test_job' do |args|
  puts "job with args #{args}" 
end

job_callback 'test_job', 'first_callback' do |args|
  puts "first callback with args #{args}"
end

job_callback 'test_job', 'second_callback' do |args|
  puts "second callback with args #{args}"
end

job_errback 'test_job', 'first_errback' do |args|
  puts "first errback with error #{args[:error]}"
end

job_errback 'test_job', 'second_errback' do |args|
  puts 'another errback'
end

Callbacks and errbacks acts as ordinary EventMachine::Deferrable callbacks and errbacks. The 'job' is the first callack, first 'job_callback' becomes second callback and so on. See lib/job_reactor/job_reactor/job_parser.rb for more information. When Node start job it calls succeed method on the 'job object' with given argument (args). This runs all callbacks sequentially. If error occurs in any callback Node calls fail method on the 'deferrable' object with the same args (plus merged :error => 'Error message).

Note, you define jobs, callbacks and errbacks in top-level scope, so the self is main object.

You can merge! additional key-value pairs to 'args' in the job to exchange information between job and it's callbacks.

include JobReactor

job 'test_job' do |args|
  args.merge!(result: 'Hello')
end

job_callback 'test_job', 'first_callback' do |args|
  puts args[:result]
  args.merge!(another_result: 'world')
end

job_callback 'test_job', 'second_callback' do |args|
  puts "#{args[:result]} #{args[:another_result]}"
end

Note, if error occurs you can't see additional arguments in job errbacks.

Another trick is JR.config[:merge_job_itself_to_args] option which is false by default. If you set this option to true you can see :job_itself key in args. The value contains many usefull information about job ('name', 'attempt', 'status', 'make_after', 'node', etc).

Feedbacks are defined as a Proc object and attached to the 'job' when it is enqueued on the application side.

success = Proc.new { |args| puts 'Success' }
error = Proc.new { |args| puts 'Error' }
JR.enqueue('my_job', {arg1: 1, arg2: 2}, {after: 100}, success, error)

This procs will be called when Node informs about success or error. The 'args' for the corresponding proc will be the same 'args' which is in the job (and it's callbacks) on the node side. So you can, for example, return any result by merging it to 'args' in the job (or it's callbacks).

Note, feedbacks are kept in memory in your application, so they disappear when you restart the application.

Job Storage

Now you can store your jobs in Redis storage ('redis_storage') (em-hiredis) or in memory ('memory_storage'). Only the first, of course, 'really' persists the jobs. You can use the last one if you don't want install Redis, don't need retry jobs and need more speed (by the way, the difference in performance is not so great - Redis is very fast). You can easily integrate your own storage. Just make it EventMachine compatible.

The default url for Redis server are:

JR.config[:hiredis_url] = "redis://127.0.0.1:6379/0"

JobReactor works asynchronously with Redis using em-hiredis library to increase the speed. Several nodes can use one Redis storage.

The informaion about jobs is saved several times during processing. This information includes:

  • id - the unique job id;
  • name - job name which 'defines' the job;
  • args - serialized arguments for the job;
  • run_at - the time when job was launched;
  • failed_at - the time when job was failed;
  • last_error - the error occured;
  • period - period (for periodic jobs);
  • defer - 'true' or 'false', flag to run job in EM.defer block;
  • status - job status ('new', 'in progress', 'queued', 'complete', 'error', 'failed', 'cancelled');
  • attempt - the number of attempt;
  • make_after - when to start job again (in seconds after last save);
  • distributor - host and port of distributor server which sent the job (used for 'feedbacks');
  • on_success - the unique id of success feedback on the distributor side;
  • on_error - the unique id of error feedback on the distributor side;

By default JobReactor deletes all completed and cancelled jobs, but you can configure it: The default options are:

JR.config[:remove_done_jobs] = true
JR.config[:remove_cancelled_jobs] = true
JR.config[:remove_failed_jobs] = false
JR.config[:retry_jobs_at_start] = true

We provide simple JR::RedisMonitor module to check the Redis storage from irb console (or from your app). We use synchronous redis gem. Connect to Redis by:

JR.config[:redis_host] = 'localhost'
JR.config[:redis_port] = 6379

See methods:

JR::RedisMonitor.jobs_for(node_name)
JR::RedisMonitor.load(job_id)
JR::RedisMonitor.destroy(job_id)
JR::RedisMonitor.destroy_all_jobs_for(node_name)

License

The MIT License - Copyright (c) 2012-2013 Anton Mishchuk