debian-mirror-gitlab/doc/development/experiment_guide/index.md
2021-03-08 18:12:59 +05:30

14 KiB

stage group info
Growth Activation To determine the technical writer assigned to the Stage/Group associated with this page, see https://about.gitlab.com/handbook/engineering/ux/technical-writing/#assignments

Experiment Guide

Experiments can be conducted by any GitLab team, most often the teams from the Growth Sub-department. Experiments are not tied to releases because they primarily target GitLab.com.

Experiments are run as an A/B test and are behind a feature flag to turn the test on or off. Based on the data the experiment generates, the team decides if the experiment had a positive impact and should be made the new default or rolled back.

Experiment tracking issue

Each experiment should have an Experiment tracking issue to track the experiment from roll-out through to cleanup/removal. Immediately after an experiment is deployed, the due date of the issue should be set (this depends on the experiment but can be up to a few weeks in the future). After the deadline, the issue needs to be resolved and either:

  • It was successful and the experiment becomes the new default.
  • It was not successful and all code related to the experiment is removed.

In either case, an outcome of the experiment should be posted to the issue with the reasoning for the decision.

Code reviews

Experiments' code quality can fail our standards for several reasons. These reasons can include not being added to the codebase for a long time, or because of fast iteration to retrieve data. However, having the experiment run (or not run) shouldn't impact GitLab availability. To avoid or identify issues, experiments are initially deployed to a small number of users. Regardless, experiments still need tests.

If, as a reviewer or maintainer, you find code that would usually fail review but is acceptable for now, mention your concerns with a note that there's no need to change the code. The author can then add a comment to this piece of code and link to the issue that resolves the experiment. If the experiment is successful and becomes part of the product, any follow up issues should be addressed.

How to create an A/B test

Implement the experiment

  1. Add the experiment to the Gitlab::Experimentation::EXPERIMENTS hash in experimentation.rb:

    EXPERIMENTS = {
      other_experiment: {
        #...
      },
      # Add your experiment here:
      signup_flow: {
        tracking_category: 'Growth::Activation::Experiment::SignUpFlow' # Used for providing the category when setting up tracking data
      }
    }.freeze
    
  2. Use the experiment in the code.

    Experiments can be performed on a subject. The subject that gets provided needs to respond to to_global_id or to_s. The resulting string is bucketed and assigned to either the control or the experimental group. It's therefore necessary to always provide the same subject for an experiment to have the same experience.

    • Use this standard for the experiment in a controller:

      Experiment run for a user:

      class ProjectController < ApplicationController
        def show
          # experiment_enabled?(:experiment_key) is also available in views and helpers
          if experiment_enabled?(:signup_flow, subject: current_user)
            # render the experiment
          else
            # render the original version
          end
        end
      end
      

      or experiment run for a namespace:

      if experiment_enabled?(:signup_flow, subject: namespace)
        # experiment code
      else
        # control code
      end
      

      When no subject is given, it falls back to a cookie that gets set and is consistent until the cookie gets deleted.

      class RegistrationController < ApplicationController
        def show
          # falls back to a cookie
          if experiment_enabled?(:signup_flow)
            # render the experiment
          else
            # render the original version
          end
        end
      end
      
    • Make the experiment available to the frontend in a controller:

      before_action do
        push_frontend_experiment(:signup_flow, subject: current_user)
      end
      

      The above checks whether the experiment is enabled and pushes the result to the frontend.

      You can check the state of the feature flag in JavaScript:

      import { isExperimentEnabled } from '~/experimentation';
      
      if ( isExperimentEnabled('signupFlow') ) {
        // ...
      }
      
    • It is also possible to run an experiment outside of the controller scope, for example in a worker:

      class SomeWorker
        def perform
          # Check if the experiment is active at all (the percentage_of_time_value > 0)
          return unless Gitlab::Experimentation.active?(:experiment_key)
      
          # Since we cannot access cookies in a worker, we need to bucket models based on a unique, unchanging attribute instead.
          # It is therefore necessery to always provide the same subject.
          if Gitlab::Experimentation.in_experiment_group?(:experiment_key, subject: user)
            # execute experimental code
          else
            # execute control code
          end
        end
      end
      

Implement the tracking events

To determine whether the experiment is a success or not, we must implement tracking events to acquire data for analyzing. We can send events to Snowplow via either the backend or frontend. Read the product intelligence guide for more details.

Track backend events

The framework provides the following helper method that is available in controllers:

before_action do
  track_experiment_event(:signup_flow, 'action', 'value', subject: current_user)
end

Which can be tested as follows:

context 'when the experiment is active and the user is in the experimental group' do
  before do
    stub_experiment(signup_flow: true)
    stub_experiment_for_subject(signup_flow: true)
  end

  it 'tracks an event', :snowplow do
    subject

    expect_snowplow_event(
      category: 'Growth::Activation::Experiment::SignUpFlow',
      action: 'action',
      value: 'value',
      label: 'experimentation_subject_id',
      property: 'experimental_group'
    )
  end
end

Track frontend events

The framework provides the following helper method that is available in controllers:

before_action do
  push_frontend_experiment(:signup_flow, subject: current_user)
  frontend_experimentation_tracking_data(:signup_flow, 'action', 'value', subject: current_user)
end

This pushes tracking data to gon.experiments and gon.tracking_data.

expect(Gon.experiments['signupFlow']).to eq(true)

expect(Gon.tracking_data).to eq(
  {
    category: 'Growth::Activation::Experiment::SignUpFlow',
    action: 'action',
    value: 'value',
    label: 'experimentation_subject_id',
    property: 'experimental_group'
  }
)

Which can then be used for tracking as follows:

import { isExperimentEnabled } from '~/lib/utils/experimentation';
import Tracking from '~/tracking';

document.addEventListener('DOMContentLoaded', () => {
  const signupFlowExperimentEnabled = isExperimentEnabled('signupFlow');

  if (signupFlowExperimentEnabled && gon.tracking_data) {
    const { category, action, ...data } = gon.tracking_data;

    Tracking.event(category, action, data);
  }
}

Which can be tested in Jest as follows:

import { withGonExperiment } from 'helpers/experimentation_helper';
import Tracking from '~/tracking';

describe('event tracking', () => {
  describe('with tracking data', () => {
    withGonExperiment('signupFlow');

    beforeEach(() => {
      jest.spyOn(Tracking, 'event').mockImplementation(() => {});

      gon.tracking_data = {
        category: 'Growth::Activation::Experiment::SignUpFlow',
        action: 'action',
        value: 'value',
        label: 'experimentation_subject_id',
        property: 'experimental_group'
      };
    });

    it('should track data', () => {
      performAction()

      expect(Tracking.event).toHaveBeenCalledWith(
        'Growth::Activation::Experiment::SignUpFlow',
        'action',
        {
          value: 'value',
          label: 'experimentation_subject_id',
          property: 'experimental_group'
        },
      );
    });
  });
});

Record experiment user

In addition to the anonymous tracking of events, we can also record which users have participated in which experiments and whether they were given the control experience or the experimental experience.

The record_experiment_user helper method is available to all controllers, and it enables you to record these experiment participants (the current user) and which experience they were given:

before_action do
  record_experiment_user(:signup_flow)
end

Subsequent calls to this method for the same experiment and the same user have no effect unless the user has gets enrolled into a different experience. This happens when we roll out the experimental experience to a greater percentage of users.

Note that this data is completely separate from the events tracking data. They are not linked together in any way.

Add context

You can add arbitrary context data in a hash which gets stored as part of the experiment user record. New calls to the record_experiment_user with newer contexts get merged deeply into the existing context.

This data can then be used by data analytics dashboards.

before_action do
  record_experiment_user(:signup_flow, foo: 42, bar: { a: 22})
  # context is { "foo" => 42, "bar" => { "a" => 22 }}
end

# Additional contexts for newer record calls are merged deeply
record_experiment_user(:signup_flow, foo: 40, bar: { b: 2 }, thor: 3)
# context becomes { "foo" => 40, "bar" => { "a" => 22, "b" => 2 }, "thor" => 3}

Record experiment conversion event

Along with the tracking of backend and frontend events and the recording of experiment participants, we can also record when a user performs the desired conversion event action. For example:

  • Experimental experience: Show an in-product nudge to see if it causes more people to sign up for trials.
  • Conversion event: The user starts a trial.

The record_experiment_conversion_event helper method is available to all controllers. It enables us to record the conversion event for the current user, regardless of whether they are in the control or experimental group:

before_action do
  record_experiment_conversion_event(:signup_flow)
end

Note that the use of this method requires that we have first recorded the user as being part of the experiment.

Enable the experiment

After all merge requests have been merged, use chatops in the appropriate channel to start the experiment for 10% of the users. The feature flag should have the name of the experiment with the _experiment_percentage suffix appended. For visibility, please also share any commands run against production in the #s_growth channel:

/chatops run feature set signup_flow_experiment_percentage 10

If you notice issues with the experiment, you can disable the experiment by removing the feature flag:

/chatops run feature delete signup_flow_experiment_percentage

Manually force the current user to be in the experiment group

You may force the application to put your current user in the experiment group. To do so add a query string parameter to the path where the experiment runs. If you do so, the experiment will work only for this request and won't work after following links or submitting forms.

For example, to forcibly enable the EXPERIMENT_KEY experiment, add force_experiment=EXPERIMENT_KEY to the URL:

https://gitlab.com/<EXPERIMENT_ENTRY_URL>?force_experiment=<EXPERIMENT_KEY>

It's possible to force the current user to be in the experiment group for <EXPERIMENT_KEY> during the browser session by using your browser's developer tools:

document.cookie = "force_experiment=<EXPERIMENT_KEY>; path=/";

Use a comma to list more than one experiment to be forced:

document.cookie = "force_experiment=<EXPERIMENT_KEY>,<ANOTHER_EXPERIMENT_KEY>; path=/";

Clear the experiments by unsetting the force_experiment cookie:

document.cookie = "force_experiment=; path=/";

Testing and test helpers

RSpec

Use the following in RSpec to mock the experiment:

context 'when the experiment is active' do
  before do
    stub_experiment(signup_flow: true)
  end

  context 'when the user is in the experimental group' do
    before do
      stub_experiment_for_subject(signup_flow: true)
    end

    it { is_expected.to do_experimental_thing }
  end

  context 'when the user is in the control group' do
    before do
      stub_experiment_for_subject(signup_flow: false)
    end

    it { is_expected.to do_control_thing }
  end
end

Jest

Use the following in Jest to mock the experiment:

import { withGonExperiment } from 'helpers/experimentation_helper';

describe('given experiment is enabled', () => {
  withGonExperiment('signupFlow');

  it('should do the experimental thing', () => {
    expect(wrapper.find('.js-some-experiment-triggered-element')).toEqual(expect.any(Element));
  });
});