ygcahyono / split

:chart_with_upwards_trend: The Rack Based AB testing framework

Split

Split is a rack based ab testing framework designed to work with Rails, Sinatra or any other rack based app.

Split is heavily inspired by the Abingo and Vanity rails ab testing plugins and Resque in its use of Redis.

Split is designed to be hacker friendly, allowing for maximum customisation and extensibility.

Backers

Support us with a monthly donation and help us continue our activities. [Become a backer]

Sponsors

Become a sponsor and get your logo on our README on Github with a link to your site. [Become a sponsor]

Requirements

Split currently requires Ruby 1.9.2 or higher. If your project requires compatibility with Ruby 1.8.x and Rails 2.3, please use v0.8.0.

Split uses redis as a datastore.

Split only supports redis 2.0 or greater.

If you're on OS X, Homebrew is the simplest way to install Redis:

brew install redis
redis-server /usr/local/etc/redis.conf

You now have a Redis daemon running on 6379.

Setup

gem install split

Rails

Adding gem 'split' to your Gemfile will autoload it when rails starts up, as long as you've configured redis it will 'just work'.

Sinatra

To configure sinatra with Split you need to enable sessions and mix in the helper methods. Add the following lines at the top of your sinatra app:

require 'split'

class MySinatraApp < Sinatra::Base
  enable :sessions
  helpers Split::Helper

  get '/' do
  ...
end

Usage

To begin your ab test use the ab_test method, naming your experiment with the first arguments. The experiment has to be predefined in the experiments configuration hash.

ab_test returns one of the alternatives, if a user has already seen that test they will get the same alternative as before, which you can use to split your code on. You can also call ab_test_result(<experiment_name>) to get chosen alternative if the user has already seen the experiment (it will return nil otherwise).

It can be used to render different templates, show different text or any other case based logic.

ab_finished is used to make a completion of an experiment, or conversion.

Example: Experiments configuration

Split.configure do |config|
  config.experiments = {
    login_button: {
      alternatives: ['/images/button1.jpg', '/images/button2.jpg']
    },
    new_user_free_points: {
      alternatives: ['100', '200', '300']
    },
    signup_page_redesign: {
      alternatives: ['first_design', 'second_design']
    }
  }
end

Example: View

<% ab_test(:login_button) do |button_file| %>
  <%= image_tag(button_file, alt: "Login!") %>
<% end %>

Example: Controller

def register_new_user
  # See what level of free points maximizes users' decision to buy replacement points.
  @starter_points = ab_test(:new_user_free_points)
end

Example: Conversion tracking (in a controller!)

def buy_new_points
  # some business logic
  ab_finished(:new_user_free_points)
end

Example: Conversion tracking (in a view)

Thanks for signing up, dude! <% ab_finished(:signup_page_redesign) %>

Statistical Validity

Split has two options for you to use to determine which alternative is the best.

The first option (default on the dashboard) uses a z test (n>30) for the difference between your control and alternative conversion rates to calculate statistical significance. This test will tell you whether an alternative is better or worse than your control, but it will not distinguish between which alternative is the best in an experiment with multiple alternatives. Split will only tell you if your experiment is 90%, 95%, or 99% significant, and this test only works if you have more than 30 participants and 5 conversions for each branch.

As per this blog post on the pitfalls of A/B testing, it is highly recommended that you determine your requisite sample size for each branch before running the experiment. Otherwise, you'll have an increased rate of false positives (experiments which show a significant effect where really there is none).

Here is a sample size calculator for your convenience.

The second option uses simulations from a beta distribution to determine the probability that the given alternative is the winner compared to all other alternatives. This option should be used when the experiment has more than just 1 control and 1 alternative. It can also be used for a simple, 2-alternative A/B test.

Extras

Weighted alternatives

Perhaps you only want to show an alternative to 10% of your visitors because it is very experimental or not yet fully load tested.

To do this you can pass a weight with each alternative in the following way:

# Experiments config:
Split.configure do |config|
  config.experiments = {
    homepage_design: {
      alternatives: [
        { name: 'Old', percent: '90' },
        { name: 'New', percent: '10' }
      ]
    }
  }
end

This will only show the new alternative to visitors 1 in 10 times.

Overriding alternatives

For development and testing, you may wish to force your app to always return an alternative. You can do this by passing it as a parameter in the url.

If you have an experiment called button_color with alternatives called red and blue used on your homepage, a url such as:

http://myawesomesite.com?ab_test[button_color]=red

will always have red buttons. This won't be stored in your session or count towards to results, unless you set the store_override configuration option.

In the event you want to disable all tests without having to know the individual experiment names, add a SPLIT_DISABLE query parameter.

http://myawesomesite.com?SPLIT_DISABLE=true

It is not required to send SPLIT_DISABLE=false to activate Split.

Starting experiments manually

By default new AB tests will be active right after deployment. In case you would like to start new test a while after the deploy, you can do it by setting the start_manually configuration option to true.

After choosing this option tests won't be started right after deploy, but after pressing the Start button in Split admin dashboard. If a test is deleted from the Split dashboard, then it can only be started after pressing the Start button whenever being re-initialized.

Reset after completion

When a user completes a test their session is reset so that they may start the test again in the future.

To stop this behaviour you can pass the following option to the ab_finished method:

ab_finished(:experiment_name, reset: false)

The user will then always see the alternative they started with.

Any old unfinished experiment key will be deleted from the user's data storage if the experiment had been removed or is over and a winner had been chosen. This allows a user to enroll into any new experiment in cases when the allow_multiple_experiments config option is set to false.

Multiple experiments at once

By default Split will avoid users participating in multiple experiments at once. This means you are less likely to skew results by adding in more variation to your tests.

To stop this behaviour and allow users to participate in multiple experiments at once set the allow_multiple_experiments config option to true like so:

Split.configure do |config|
  config.allow_multiple_experiments = true
end

This will allow the user to participate in any number of experiments and belong to any alternative in each experiment. This has the possible downside of a variation in one experiment influencing the outcome of another.

To address this, setting the allow_multiple_experiments config option to 'control' like so:

Split.configure do |config|
  config.allow_multiple_experiments = 'control'
end

For this to work, each and every experiment you define must have an alternative named 'control'. This will allow the user to participate in multiple experiments as long as the user belongs to the alternative 'control' in each experiment. As soon as the user belongs to an alternative named something other than 'control' the user may not participate in any more experiments. Calling ab_test() will always return the first alternative without adding the user to that experiment.

Experiment Persistence

Split comes with three built-in persistence adapters for storing users and the alternatives they've been given for each experiment.

By default Split will store the tests for each user in the session.

You can optionally configure Split to use a cookie, Redis, or any custom adapter of your choosing.

Cookies

Split.configure do |config|
  config.persistence = :cookie
end

By default, cookies will expire in 1 year. To change that, set the persistence_cookie_length in the configuration (unit of time in seconds).

Split.configure do |config|
  config.persistence = :cookie
  config.persistence_cookie_length = 2592000 # 30 days
end

Note: Using cookies depends on ActionDispatch::Cookies or any identical API

Redis

Using Redis will allow ab_users to persist across sessions or machines.

Split.configure do |config|
  config.persistence = Split::Persistence::RedisAdapter.with_config(lookup_by: -> (context) { context.current_user_id })
  # Equivalent
  # config.persistence = Split::Persistence::RedisAdapter.with_config(lookup_by: :current_user_id)
end

Options:

• lookup_by: method to invoke per request for uniquely identifying ab_users (mandatory configuration)
• namespace: separate namespace to store these persisted values (default "persistence")
• expire_seconds: sets TTL for user key. (if a user is in multiple experiments most recent update will reset TTL for all their assignments)

Dual Adapter

The Dual Adapter allows the use of different persistence adapters for logged-in and logged-out users. A common use case is to use Redis for logged-in users and Cookies for logged-out users.

cookie_adapter = Split::Persistence::CookieAdapter
redis_adapter = Split::Persistence::RedisAdapter.with_config(
    lookup_by: -> (context) { context.send(:current_user).try(:id) },
    expire_seconds: 2592000)

Split.configure do |config|
  config.persistence = Split::Persistence::DualAdapter.with_config(
      logged_in: -> (context) { !context.send(:current_user).try(:id).nil? },
      logged_in_adapter: redis_adapter,
      logged_out_adapter: cookie_adapter)
  config.persistence_cookie_length = 2592000 # 30 days
end

Custom Adapter

Your custom adapter needs to implement the same API as existing adapters. See Split::Persistence::CookieAdapter or Split::Persistence::SessionAdapter for a starting point.

Split.configure do |config|
  config.persistence = YourCustomAdapterClass
end

Trial Event Hooks

You can define methods that will be called at the same time as experiment alternative participation and goal completion.

For example:

Split.configure do |config|
  config.on_trial  = :log_trial # run on every trial
  config.on_trial_choose   = :log_trial_choose # run on trials with new users only
  config.on_trial_complete = :log_trial_complete
end

Set these attributes to a method name available in the same context as the ab_test method. These methods should accept one argument, a Trial instance.

def log_trial(trial)
  logger.info "experiment=%s alternative=%s user=%s" %
    [ trial.experiment.name, trial.alternative, current_user.id ]
end

def log_trial_choose(trial)
  logger.info "[new user] experiment=%s alternative=%s user=%s" %
    [ trial.experiment.name, trial.alternative, current_user.id ]
end

def log_trial_complete(trial)
  logger.info "experiment=%s alternative=%s user=%s complete=true" %
    [ trial.experiment.name, trial.alternative, current_user.id ]
end

Views

If you are running ab_test from a view, you must define your event hook callback as a helper_method in the controller:

helper_method :log_trial_choose

def log_trial_choose(trial)
  logger.info "experiment=%s alternative=%s user=%s" %
    [ trial.experiment.name, trial.alternative, current_user.id ]
end

Experiment Hooks

You can assign a proc that will be called when an experiment is reset or deleted. You can use these hooks to call methods within your application to keep data related to experiments in sync with Split.

For example:

Split.configure do |config|
  # after experiment reset or deleted
  config.on_experiment_reset  = -> (example) { # Do something on reset }
  config.on_experiment_delete = -> (experiment) { # Do something else on delete }
  # before experiment reset or deleted
  config.on_before_experiment_reset  = -> (example) { # Do something on reset }
  config.on_before_experiment_delete = -> (experiment) { # Do something else on delete }
end

Web Interface

Split comes with a Sinatra-based front end to get an overview of how your experiments are doing.

If you are running Rails 2: You can mount this inside your app using Rack::URLMap in your config.ru

require 'split/dashboard'

run Rack::URLMap.new \
  "/"       => Your::App.new,
  "/split" => Split::Dashboard.new

However, if you are using Rails 3: You can mount this inside your app routes by first adding this to the Gemfile:

gem 'split', require: 'split/dashboard'

Then adding this to config/routes.rb

mount Split::Dashboard, at: 'split'

You may want to password protect that page, you can do so with Rack::Auth::Basic (in your split initializer file)

# Rails apps or apps that already depend on activesupport
Split::Dashboard.use Rack::Auth::Basic do |username, password|
  # Protect against timing attacks:
  # - Use & (do not use &&) so that it doesn't short circuit.
  # - Use `variable_size_secure_compare` to stop length information leaking
  ActiveSupport::SecurityUtils.variable_size_secure_compare(username, ENV["SPLIT_USERNAME"]) &
    ActiveSupport::SecurityUtils.variable_size_secure_compare(password, ENV["SPLIT_PASSWORD"])
end

# Apps without activesupport
Split::Dashboard.use Rack::Auth::Basic do |username, password|
  # Protect against timing attacks:
  # - Use & (do not use &&) so that it doesn't short circuit.
  # - Use digests to stop length information leaking
  Rack::Utils.secure_compare(::Digest::SHA256.hexdigest(username), ::Digest::SHA256.hexdigest(ENV["SPLIT_USERNAME"])) &
    Rack::Utils.secure_compare(::Digest::SHA256.hexdigest(password), ::Digest::SHA256.hexdigest(ENV["SPLIT_PASSWORD"]))
end

You can even use Devise or any other Warden-based authentication method to authorize users. Just replace mount Split::Dashboard, :at => 'split' in config/routes.rb with the following:

match "/split" => Split::Dashboard, anchor: false, via: [:get, :post, :delete], constraints: -> (request) do
  request.env['warden'].authenticated? # are we authenticated?
  request.env['warden'].authenticate! # authenticate if not already
  # or even check any other condition such as request.env['warden'].user.is_admin?
end

More information on this here

Screenshot

Configuration

You can override the default configuration options of Split like so:

Split.configure do |config|
  config.db_failover = true # handle redis errors gracefully
  config.db_failover_on_db_error = -> (error) { Rails.logger.error(error.message) }
  config.allow_multiple_experiments = true
  config.enabled = true
  config.persistence = Split::Persistence::SessionAdapter
  #config.start_manually = false ## new test will have to be started manually from the admin panel. default false
  config.include_rails_helper = true
  config.redis = "redis://custom.redis.url:6380"
end

Split looks for the Redis host in the environment variable REDIS_URL then defaults to redis://localhost:6379 if not specified by configure block.

On platforms like Heroku, Split will use the value of REDIS_PROVIDER to determine which env variable key to use when retrieving the host config. This defaults to REDIS_URL.

Filtering

In most scenarios you don't want to have AB-Testing enabled for web spiders, robots or special groups of users. Split provides functionality to filter this based on a predefined, extensible list of bots, IP-lists or custom exclude logic.

Split.configure do |config|
  # bot config
  config.robot_regex = /my_custom_robot_regex/ # or
  config.bots['newbot'] = "Description for bot with 'newbot' user agent, which will be added to config.robot_regex for exclusion"

  # IP config
  config.ignore_ip_addresses << '81.19.48.130' # or regex: /81\.19\.48\.[0-9]+/

  # or provide your own filter functionality, the default is proc{ |request| is_robot? || is_ignored_ip_address? }
  config.ignore_filter = -> (request) { CustomExcludeLogic.excludes?(request) }
end

Experiment configuration

Instead of providing the experiment options inline, you can store them in a hash. This hash can control your experiment's alternatives, weights, algorithm and if the experiment resets once finished:

Split.configure do |config|
  config.experiments = {
    my_first_experiment: {
      alternatives: ["a", "b"],
      resettable: false
    },
    :my_second_experiment => {
      algorithm: 'Split::Algorithms::Whiplash',
      alternatives: [
        { name: "a", percent: 67 },
        { name: "b", percent: 33 }
      ]
    }
  }
end

You can also store your experiments in a YAML file:

Split.configure do |config|
  config.experiments = YAML.load_file "config/experiments.yml"
end

You can then define the YAML file like:

my_first_experiment:
  alternatives:
    - a
    - b
my_second_experiment:
  alternatives:
    - name: a
      percent: 67
    - name: b
      percent: 33
  resettable: false

This simplifies the calls from your code:

ab_test(:my_first_experiment)

and:

ab_finished(:my_first_experiment)

You can also add meta data for each experiment, very useful when you need more than an alternative name to change behaviour:

Split.configure do |config|
  config.experiments = {
    my_first_experiment: {
      alternatives: ["a", "b"],
      metadata: {
        "a" => {"text" => "Have a fantastic day"},
        "b" => {"text" => "Don't get hit by a bus"}
      }
    }
  }
end

my_first_experiment:
  alternatives:
    - a
    - b
  metadata:
    a:
      text: "Have a fantastic day"
    b:
      text: "Don't get hit by a bus"

This allows for some advanced experiment configuration using methods like:

trial.alternative.name # => "a"

trial.metadata['text'] # => "Have a fantastic day"

or in views:

<% ab_test("my_first_experiment") do |alternative, meta| %>
  <%= alternative %>
  <small><%= meta['text'] %></small>
<% end %>

Scores

You can also track generic values alongside each alternatives. You can use the configuration hash to do this with the :scores option:

Split.configure do |config|
  config.experiments = {
    my_first_experiment: {
      alternatives: ["alt_one", "alt_two"],
      scores: ["score_one", "score_two"]
    },
    my_second_experiment: {
      alternatives: ["alt_three", "alt_four"],
      scores: ["score_one", "score_three"]
    }
  }
end

Consider the case when you want to track specific value to each of the ongoing experiments, you can do so by calling:

ab_score('score_one', 100)

in a session where the user enrolls to my_first_experiment and my_second_experiment simultaneously, each alternative chosen on both experiment will then has its score score_one incremented by 100. These scores are also being displayed on split dashboard.

You can also give a score but only add the value at later point:

# in a controller
ab_add_delayed_score('score_one', 'product_123', 100, 60)

# anywhere in the code
ab_apply_delayed_score('score_one', 'product_123')

ab_add_delayed_score will keep track of the score and current running experiments (without actually adding the value to the total score yet), and ab_apply_delayed_score will then add the value to each tracked alternatives. From left to right, the parameters for ab_add_delayed_score are:

• 'score_one' - String, the name of the score
• 'product_123' - String, identifier for the delayed score
• 100 - Integer, the value of the score (optional, default: 1)
• 60 - Integer, the TTL in seconds for the delayed score before it's deleted from database and can't be applied afterwards. (optional, default: 1 day)

Only the first 2 parameters are used for ab_apply_delayed_score

It is also possible to directly add a score to an alternative without being in a web session:

Split::Helper.ab_score_alternative(:my_first_experiment, "alt_one", "score_one", 100)

DB failover solution

Due to the fact that Redis has no automatic failover mechanism, it's possible to switch on the db_failover config option, so that ab_test and ab_finished will not crash in case of a db failure. ab_test always delivers alternative A (the first one) in that case.

It's also possible to set a db_failover_on_db_error callback (proc) for example to log these errors via Rails.logger.

Redis

You may want to change the Redis host and port Split connects to, or set various other options at startup.

Split has a redis setter which can be given a string or a Redis object. This means if you're already using Redis in your app, Split can re-use the existing connection.

String: Split.redis = 'redis://localhost:6379'

Redis: Split.redis = $redis

For our rails app we have a config/initializers/split.rb file where we load config/split.yml by hand and set the Redis information appropriately.

Here's our config/split.yml:

development: redis://localhost:6379
test: redis://localhost:6379
staging: redis://redis1.example.com:6379
fi: redis://localhost:6379
production: redis://redis1.example.com:6379

And our initializer:

split_config = YAML.load_file(Rails.root.join('config', 'split.yml'))
Split.redis = split_config[Rails.env]

Namespaces

If you're running multiple, separate instances of Split you may want to namespace the keyspaces so they do not overlap. This is not unlike the approach taken by many memcached clients.

This feature can be provided by the redis-namespace library. To configure Split to use Redis::Namespace, do the following:

1. Add redis-namespace to your Gemfile:

gem 'redis-namespace'

2. Configure Split.redis to use a Redis::Namespace instance (possible in an intializer):

redis = Redis.new(url: ENV['REDIS_URL']) # or whatever config you want
Split.redis = Redis::Namespace.new(:your_namespace, redis: redis)

Outside of a Web Session

Split provides the Helper module to facilitate running experiments inside web sessions.

Alternatively, you can access the underlying Trial, Experiment and Alternative objects to conduct experiments that are not tied to a web session.

# create a new experiment
experiment = Split::ExperimentCatalog.find_or_create('color', 'red', 'blue')
# create a new trial
trial = Split::Trial.new(:experiment => experiment)
# run trial
trial.choose!
# get the result, returns either red or blue
trial.alternative.name

# if the goal has been achieved, increment the successful completions for this alternative.
if goal_achieved?
  trial.complete!
end

Algorithms

By default, Split ships with Split::Algorithms::WeightedSample that randomly selects from possible alternatives for a traditional a/b test. It is possible to specify static weights to favor certain alternatives.

Split::Algorithms::Whiplash is an implementation of a multi-armed bandit algorithm. This algorithm will automatically weight the alternatives based on their relative performance, choosing the better-performing ones more often as trials are completed.

Users may also write their own algorithms. The default algorithm may be specified globally in the configuration file, or on a per experiment basis using the experiments hash of the configuration file.

To change the algorithm globally for all experiments, use the following in your initializer:

Split.configure do |config|
  config.algorithm = Split::Algorithms::Whiplash
end

Extensions

• Split::Export - easily export ab test data out of Split
• Split::Analytics - push test data to google analytics
• Split::Mongoid - store experiment data in mongoid (still uses redis)
• Split::Cacheable - automatically create cache buckets per test
• Split::Counters - add counters per experiment and alternative
• Split::Cli - a CLI to trigger Split A/B tests

Screencast

Ryan bates has produced an excellent 10 minute screencast about split on the Railscasts site: A/B Testing with Split

Blogposts

• A/B Testing with Split in Ruby on Rails
• Recipe: A/B testing with KISSMetrics and the split gem
• Rails A/B testing with Split on Heroku

Contributors

Over 70 different people have contributed to the project, you can see them all here: https://github.com/splitrb/split/graphs/contributors

Development

Run the tests like this:

# Start a Redis server in another tab.
redis-server

bundle
rake spec

Source hosted at GitHub.

Report Issues/Feature requests on GitHub Issues.

Discussion at Google Groups.

Note on Patches/Pull Requests

• Fork the project.
• Make your feature addition or bug fix.
• Add tests for it. This is important so I don't break it in a future version unintentionally.
• Add documentation if necessary.
• Commit, do not mess with rakefile, version, or history. (if you want to have your own version, that is fine but bump version in a commit by itself I can ignore when I pull)
• Send a pull request. Bonus points for topic branches.

Copyright

Copyright (c) 2016 Andrew Nesbitt. See LICENSE for details.