Awesome

Paxuby

Paxos algorithm in Ruby

Introduction

The Paxos algorithm is a cornerstone in building distributed, fault-tolerant storage systems. Notable real world applications include Chubby, Google's lock managing system that serves the Google Filesystem(GFS) and Gaios, Microsoft's replicated datastore. Regardless of its backend, under the assumption that only less than half of a system's instances may encouter failures, Paxos guarantees that:

• service continues to be available without performance compromises
• data remains consistent at every single moment (compare Eventual consistency)

Besides the above guarantess, this particular implementation aims at providing:

• a convenient interface to add Paxos support for arbitrary backends, possibly a SQL database or Redis etc. (a feature that is theoretically given but less seen in actually implementations)
• easier maintenance and perhaps higher educational value by being written in a high level language Ruby, while sustaining as much performance as possible (please see the performance section below)

Usage

Assuming you are using the default key-value backend,

• Set up 'config.yaml' to list the IPs of your cluster, for example

  addrs:
    - 127.0.0.1:6660
    - 127.0.0.1:6661
    - 127.0.0.1:6662
  

• Run ruby main.rb in each machine

• Interract with Paxos using the built-in client:

  irb> require './app/client'
  irb> client = App::Client.new '127.0.0.1:6660'
  irb> client.get 'some_key'
    => 'nil'
  irb> client.set 'some_key', 'the_value'
  irb> client.get 'some_key'
    => 'the_value'
  

• Enjoy the fault-tolerance provided by Paxos by stopping one of the processes perhaps using ctrl-C.

Interface between paxuby and backends

The main interface lies in the App module. Custom backends can be supported by implementing:

• the class App::Command which assigns the instance variable @err_msg when there's an error parsing a command string
• the module_function App.execute, which is expected to return a Paxos::App::Result object.

A straight forward example can be found in './app/app.rb'.

Testing and Profiling

• Run bacon test/test.rb to run all tests. The logs of the test can be displayed by ps aux | grep ruby; cat tmp/*log; echo 'select * from paxos;' | sqlite3 '127.0.0.1:6660paxos.db'
• Run ruby test/perf.rb to test performance under normal situations where the leader does not change, but with replicas free to fail. The result printed out is the response time of a client request vesus its frequency. In addition, by default we also use perftools.rb to profile our code.
• Run ruby test/leader_promotion_perf.rb to test performance for the time spent when a leader is dead, and the system automatically elects a new leader.

Performance

The results of ruby test/perf.rb for the response time is 7.5ms for the 50th percentil, 24ms for the 90th, but 600ms for the 99th. The below charts show the distribution (note that we omited the long tail which is actually important in real world applications) and the results of gperftools for the leader and replicas (wallclock time):

Config options

We read the configs from 'config.yaml' in YAML format:

addrs:
  - 127.0.0.1:6660
  - 127.0.0.1:6661
  - 127.0.0.1:6662
heartbeat_timeout: 0.75 # time for replicas to check whether leader is alive, in seconds
perftools: false # enable gperftools profiling

License

Copyright 2013 Awaw Fumin awawfumin@gmail.com
Licensed under GNU Affero General Public License v3

Todos

• implement snapshots
• improve performance according to gperftools results