Implementing an A/B/n experiment

Introduction

Experiments in GitLab are tightly coupled with the concepts provided by Feature flags in development of GitLab. You're strongly encouraged to read and understand the Feature flags in development of GitLab portion of the documentation before considering running experiments. Experiments add additional concepts which may seem confusing or advanced without understanding the underpinnings of how GitLab uses feature flags in development. One concept: experiments can be run with multiple variants, which are sometimes referred to as A/B/n tests.

We use the gitlab-experiment gem, sometimes referred to as GLEX, to run our experiments. The gem exists in a separate repository so it can be shared across any GitLab property that uses Ruby. You should feel comfortable reading the documentation on that project if you want to dig into more advanced topics or open issues. Be aware that the documentation there reflects what's in the main branch and may not be the same as the version being used within GitLab.

Glossary of terms

To ensure a shared language, you should understand these fundamental terms we use when communicating about experiments:

  • experiment: Any deviation of code paths we want to run at some times, but not others.
  • context: A consistent experience we provide in an experiment.
  • control: The default, or "original" code path.
  • candidate: Defines an experiment with only one code path.
  • variant(s): Defines an experiment with multiple code paths.
  • behaviors: Used to reference all possible code paths of an experiment, including the control.

Implementing an experiment

Examples

Start by generating a feature flag using the bin/feature-flag command as you normally would for a development feature flag, making sure to use experiment for the type. For the sake of documentation let's name our feature flag (and experiment) "pill_color".

bin/feature-flag pill_color -t experiment

After you generate the desired feature flag, you can immediately implement an experiment in code. An experiment implementation can be as simple as:

experiment(:pill_color, actor: current_user) do |e|
  e.control { 'control' }
  e.variant(:red) { 'red' }
  e.variant(:blue) { 'blue' }
end

When this code executes, the experiment is run, a variant is assigned, and (if within a controller or view) a window.gl.experiments.pill_color object will be available in the client layer, with details like:

  • The assigned variant.
  • The context key for client tracking events.

In addition, when an experiment runs, an event is tracked for the experiment :assignment. We cover more about events, tracking, and the client layer later.

In local development, you can make the experiment active by using the feature flag interface. You can also target specific cases by providing the relevant experiment to the call to enable the feature flag:

# Enable for everyone
Feature.enable(:pill_color)

# Get the `experiment` method -- already available in controllers, views, and mailers.
include Gitlab::Experiment::Dsl
# Enable for only the first user
Feature.enable(:pill_color, experiment(:pill_color, actor: User.first))

To roll out your experiment feature flag on an environment, run the following command using ChatOps (which is covered in more depth in the Feature flags in development of GitLab documentation). This command creates a scenario where half of everyone who encounters the experiment would be assigned the control, 25% would be assigned the red variant, and 25% would be assigned the blue variant:

/chatops run feature set pill_color 50 --actors

For an even distribution in this example, change the command to set it to 66% instead of 50.

NOTE: To immediately stop running an experiment, use the /chatops run feature set pill_color false command.

WARNING: We strongly recommend using the --actors flag when using the ChatOps commands, as anything else may give odd behaviors due to how the caching of variant assignment is handled.

We can also implement this experiment in a HAML file with HTML wrappings:

#cta-interface
  - experiment(:pill_color, actor: current_user) do |e|
    - e.control do
      .pill-button control
    - e.variant(:red) do
      .pill-button.red red
    - e.variant(:blue) do
      .pill-button.blue blue

The importance of context

In our previous example experiment, our context (this is an important term) is a hash that's set to { actor: current_user }. Context must be unique based on how you want to run your experiment, and should be understood at a lower level.

It's expected, and recommended, that you use some of these contexts to simplify reporting:

  • { actor: current_user }: Assigns a variant and is "sticky" to each user (or "client" if current_user is nil) who enters the experiment.
  • { project: project }: Assigns a variant and is "sticky" to the project currently being viewed. If running your experiment is more useful when viewing a project, rather than when a specific user is viewing any project, consider this approach.
  • { group: group }: Similar to the project example, but applies to a wider scope of projects and users.
  • { actor: current_user, project: project }: Assigns a variant and is "sticky" to the user who is viewing the given project. This creates a different variant assignment possibility for every project that current_user views. Understand this can create a large cache size if an experiment like this in a highly trafficked part of the application.
  • { wday: Time.current.wday }: Assigns a variant based on the current day of the week. In this example, it would consistently assign one variant on Friday, and a potentially different variant on Saturday.

Context is critical to how you define and report on your experiment. It's usually the most important aspect of how you choose to implement your experiment, so consider it carefully, and discuss it with the wider team if needed. Also, take into account that the context you choose affects our cache size.

After the above examples, we can state the general case: given a specific and consistent context, we can provide a consistent experience and track events for that experience. To dive a bit deeper into the implementation details: a context key is generated from the context that's provided. Use this context key to:

  • Determine the assigned variant.
  • Identify events tracked against that context key.

We can think about this as the experience that we've rendered, which is both dictated and tracked by the context key. The context key is used to track the interaction and results of the experience we've rendered to that context key. These concepts are somewhat abstract and hard to understand initially, but this approach enables us to communicate about experiments as something that's wider than just user behavior.

NOTE: Using actor: utilizes cookies if the current_user is nil. If you don't need cookies though - meaning that the exposed functionality would only be visible to signed in users - { user: current_user } would be just as effective.

WARNING: The caching of variant assignment is done by using this context, and so consider your impact on the cache size when defining your experiment. If you use { time: Time.current } you would be inflating the cache size every time the experiment is run. Not only that, your experiment would not be "sticky" and events wouldn't be resolvable.

Advanced experimentation

There are two ways to implement an experiment:

  1. The simple experiment style described previously.
  2. A more advanced style where an experiment class is provided.

The advanced style is handled by naming convention, and works similar to what you would expect in Rails.

To generate a custom experiment class that can override the defaults in ApplicationExperiment use the Rails generator:

rails generate gitlab:experiment pill_color control red blue

This generates an experiment class in app/experiments/pill_color_experiment.rb with the behaviors we've provided to the generator. Here's an example of how that class would look after migrating our previous example into it:

class PillColorExperiment < ApplicationExperiment
  control { 'control' }
  variant(:red) { 'red' }
  variant(:blue) { 'blue' }
end

We can now simplify where we run our experiment to the following call, instead of providing the block we were initially providing, by explicitly calling run:

experiment(:pill_color, actor: current_user).run

The behaviors we defined in our experiment class represent the default implementation. You can still use the block syntax to override these behaviors however, so the following would also be valid:

experiment(:pill_color, actor: current_user) do |e|
  e.control { '<strong>control</strong>' }
end

NOTE: When passing a block to the experiment method, it is implicitly invoked as if run has been called.

Segmentation rules

You can use runtime segmentation rules to, for instance, segment contexts into a specific variant. The segment method is a callback (like before_action) and so allows providing a block or method name.

In this example, any user named 'Richard' would always be assigned the red variant, and any account older than 2 weeks old would be assigned the blue variant:

class PillColorExperiment < ApplicationExperiment
  # ...registered behaviors

  segment(variant: :red) { context.actor.first_name == 'Richard' }
  segment :old_account?, variant: :blue

  private

  def old_account?
    context.actor.created_at < 2.weeks.ago
  end
end

When an experiment runs, the segmentation rules are executed in the order they're defined. The first segmentation rule to produce a truthy result assigns the variant.

In our example, any user named 'Richard', regardless of account age, will always be assigned the red variant. If you want the opposite logic, flip the order.

NOTE: Keep in mind when defining segmentation rules: after a truthy result, the remaining segmentation rules are skipped to achieve optimal performance.

Exclusion rules

Exclusion rules are similar to segmentation rules, but are intended to determine if a context should even be considered as something we should include in the experiment and track events toward. Exclusion means we don't care about the events in relation to the given context.

These examples exclude all users named 'Richard', and any account older than 2 weeks old. Not only are they given the control behavior - which could be nothing - but no events are tracked in these cases as well.

class PillColorExperiment < ApplicationExperiment
  # ...registered behaviors

  exclude :old_account?, ->{ context.actor.first_name == 'Richard' }

  private

  def old_account?
    context.actor.created_at < 2.weeks.ago
  end
end

You may also need to check exclusion in custom tracking logic by calling should_track?:

class PillColorExperiment < ApplicationExperiment
  # ...registered behaviors

  def expensive_tracking_logic
    return unless should_track?

    track(:my_event, value: expensive_method_call)
  end
end

Tracking events

One of the most important aspects of experiments is gathering data and reporting on it. You can use the track method to track events across an experimental implementation. You can track events consistently to an experiment if you provide the same context between calls to your experiment. If you do not yet understand context, you should read about contexts now.

We can assume we run the experiment in one or a few places, but track events potentially in many places. The tracking call remains the same, with the arguments you would normally use when tracking events using snowplow. The easiest example of tracking an event in Ruby would be:

experiment(:pill_color, actor: current_user).track(:clicked)

When you run an experiment with any of the examples so far, an :assignment event is tracked automatically by default. All events that are tracked from an experiment have a special experiment context added to the event. This can be used - typically by the data team - to create a connection between the events on a given experiment.

If our current user hasn't encountered the experiment yet (meaning where the experiment is run), and we track an event for them, they are assigned a variant and see that variant if they ever encountered the experiment later, when an :assignment event would be tracked at that time for them.

NOTE: GitLab tries to be sensitive and respectful of our customers regarding tracking, so our experimentation library allows us to implement an experiment without ever tracking identifying IDs. It's not always possible, though, based on experiment reporting requirements. You may be asked from time to time to track a specific record ID in experiments. The approach is largely up to the PM and engineer creating the implementation. No recommendations are provided here at this time.

Testing with RSpec

In the course of working with experiments, you'll probably want to utilize the RSpec tooling that's built in. This happens automatically for files in spec/experiments, but for other files and specs you want to include it in, you can specify the :experiment type:

it "tests experiments nicely", :experiment do
end

Stub helpers

You can stub experiments using stub_experiments. Pass it a hash using experiment names as the keys, and the variants you want each to resolve to, as the values:

# Ensures the experiments named `:example` & `:example2` are both "enabled" and
# that each will resolve to the given variant (`:my_variant` and `:control`
# respectively).
stub_experiments(example: :my_variant, example2: :control)

experiment(:example) do |e|
  e.enabled? # => true
  e.assigned.name # => 'my_variant'
end

experiment(:example2) do |e|
  e.enabled? # => true
  e.assigned.name # => 'control'
end

Exclusion, segmentation, and behavior matchers

You can also test things like the registered behaviors, the exclusions, and segmentations using the matchers.

class ExampleExperiment < ApplicationExperiment
  control { }
  candidate { '_candidate_' }
    
  exclude { context.actor.first_name == 'Richard' }
  segment(variant: :candidate) { context.actor.username == 'jejacks0n' }
end

excluded = double(username: 'rdiggitty', first_name: 'Richard')
segmented = double(username: 'jejacks0n', first_name: 'Jeremy')

# register_behavior matcher
expect(experiment(:example)).to register_behavior(:control)
expect(experiment(:example)).to register_behavior(:candidate).with('_candidate_')

# exclude matcher
expect(experiment(:example)).to exclude(actor: excluded)
expect(experiment(:example)).not_to exclude(actor: segmented)

# segment matcher
expect(experiment(:example)).to segment(actor: segmented).into(:candidate)
expect(experiment(:example)).not_to segment(actor: excluded)

Tracking matcher

Tracking events is a major aspect of experimentation. We try to provide a flexible way to ensure your tracking calls are covered.

You can do this on the instance level or at an "any instance" level:

subject = experiment(:example)

expect(subject).to track(:my_event)

subject.track(:my_event)

You can use the on_next_instance chain method to specify that it will happen on the next instance of the experiment. This helps you if you're calling experiment(:example).track downstream:

expect(experiment(:example)).to track(:my_event).on_next_instance

experiment(:example).track(:my_event)

A full example of the methods you can chain onto the track matcher:

expect(experiment(:example)).to track(:my_event, value: 1, property: '_property_')
  .on_next_instance
  .with_context(foo: :bar)
  .for(:variant_name)

experiment(:example, :variant_name, foo: :bar).track(:my_event, value: 1, property: '_property_')

Experiments in the client layer

Any experiment that's been run in the request lifecycle surfaces in window.gl.experiments, and matches this schema so it can be used when resolving experimentation in the client layer.

Given that we've defined a class for our experiment, and have defined the variants for it, we can publish that experiment in a couple ways.

The first way is simply by running the experiment. Assuming the experiment has been run, it will surface in the client layer without having to do anything special.

The second way doesn't run the experiment and is intended to be used if the experiment only needs to surface in the client layer. To accomplish this we can simply .publish the experiment. This won't run any logic, but does surface the experiment details in the client layer so they can be utilized there.

An example might be to publish an experiment in a before_action in a controller. Assuming we've defined the PillColorExperiment class, like we have above, we can surface it to the client by publishing it instead of running it:

before_action -> { experiment(:pill_color).publish }, only: [:show]

You can then see this surface in the JavaScript console:

window.gl.experiments // => { pill_color: { excluded: false, experiment: "pill_color", key: "ca63ac02", variant: "candidate" } } 

Using experiments in Vue

With the gitlab-experiment component, you can define slots that match the name of the variants pushed to window.gl.experiments.

We can make use of the named slots in the Vue component, that match the behaviors defined in :

<script>
import GitlabExperiment from '~/experimentation/components/gitlab_experiment.vue';

export default {
  components: { GitlabExperiment }
}
</script>

<template>
  <gitlab-experiment name="pill_color">
    <template #control>
      <button class="bg-default">Click default button</button>
    </template>

    <template #red>
      <button class="bg-red">Click red button</button>
    </template>

    <template #blue>
      <button class="bg-blue">Click blue button</button>
    </template>
  </gitlab-experiment>
</template>

NOTE: When there is no experiment data in the window.gl.experiments object for the given experiment name, the control slot will be used, if it exists.

Test with Jest

Stub Helpers

You can stub experiments using the stubExperiments helper defined in spec/frontend/__helpers__/experimentation_helper.js.

import { stubExperiments } from 'helpers/experimentation_helper';
import { getExperimentData } from '~/experimentation/utils';

describe('when my_experiment is enabled', () => {
  beforeEach(() => {
    stubExperiments({ my_experiment: 'candidate' });
  });

  it('sets the correct data', () => {
    expect(getExperimentData('my_experiment')).toEqual({ experiment: 'my_experiment', variant: 'candidate' });
  });
});

NOTE: This method of stubbing in Jest specs will not automatically un-stub itself at the end of the test. We merge our stubbed experiment in with all the other global data in window.gl. If you need to remove the stubbed experiment(s) after your test or ensure a clean global object before your test, you'll need to manage the global object directly yourself:

describe('tests that care about global state', () => {
  const originalObjects = [];

  beforeEach(() => {
    // For backwards compatibility for now, we're using both window.gon & window.gl
    originalObjects.push(window.gon, window.gl);
  });

  afterEach(() => {
    [window.gon, window.gl] = originalObjects;
  });

  it('stubs experiment in fresh global state', () => {
    stubExperiment({ my_experiment: 'candidate' });
    // ...
  });
})

Notes on feature flags

NOTE: We use the terms "enabled" and "disabled" here, even though it's against our documentation style guide recommendations because these are the terms that the feature flag documentation uses.

You may already be familiar with the concept of feature flags in GitLab, but using feature flags in experiments is a bit different. While in general terms, a feature flag is viewed as being either on or off, this isn't accurate for experiments.

Generally, off means that when we ask if a feature flag is enabled, it will always return false, and on means that it will always return true. An interim state, considered conditional, also exists. We take advantage of this trinary state of feature flags. To understand this conditional aspect: consider that either of these settings puts a feature flag into this state:

  • Setting a percentage_of_actors of any percent greater than 0%.
  • Enabling it for a single user or group.

Conditional means that it returns true in some situations, but not all situations.

When a feature flag is disabled (meaning the state is off), the experiment is considered inactive. You can visualize this in the decision tree diagram as reaching the first Running? node, and traversing the negative path.

When a feature flag is rolled out to a percentage_of_actors or similar (meaning the state is conditional) the experiment is considered to be running where sometimes the control is assigned, and sometimes the candidate is assigned. We don't refer to this as being enabled, because that's a confusing and overloaded term here. In the experiment terms, our experiment is running, and the feature flag is conditional.

When a feature flag is enabled (meaning the state is on), the candidate will always be assigned.

We should try to be consistent with our terms, and so for experiments, we have an inactive experiment until we set the feature flag to conditional. After which, our experiment is then considered running. If you choose to "enable" your feature flag, you should consider the experiment to be resolved, because everyone is assigned the candidate unless they've opted out of experimentation.

As of GitLab 13.10, work is being done to improve this process and how we communicate about it.