ADR-73: A/B testing

More details about this document
Latest published version:
GitHub decentraland/adr (pull requests, new issue, open issues)
Edit this documentation:
GitHub View commits View commits on

Context and Problem Statement

Given the number of new features we have, we need a way to measure the impact that each one has on the use of our products. A/B testing is what we use to decide between two or more versions of a new feature and/or to compare it with the current one. This ADR describes how we measure A/B testings in our products.

Main objectives

Current implementation

To setup an A/B testing we use Feature Flags as an assignation tool and Segment as a measurement tool.

Feature Flags: Setting up a new A/B testing

Using this standard, an A/B is defined as a feature flag that ends with _variant with each of the possible variants defined as a variant and the activation defined as your application required.

Please note that if you need to test a new variant against the current one it is preferable to use a variant with two options (enabled and disabled) instead of the activation because otherwise, you won't be able to differentiate (and measure) when your A/B is active or just should show the current version.

Creating a new flag

Just create a new feature that ends _variant

Create new flag

Defining your activation

You can define your activation as your project needs, but in most of the cases you will need one of this two options:

applicationHostname allows you to keep your test active on any testing environment and when its active on production

Activation by host

If you also add gradualRolloutSessionId you can rollout your test gradually

Activation rollout

Defining your variants

In the VARIANT tab you can define as many cases as you need and control the percentage of users that will get each variant.

Add variant
List variants

Please note that if you define and activation with rollout and enabled/disabled variants you will split your users as follow:

For a total of 1.000 only half of them will count as testing users and half of those will see the enabled options, that means that 750 users will see the old version but only 250 will be considered in the result of the A/B testing. If your objective is to test 500 users with enabled and 500 users with disabled remove the rollout configuration.

Once your A/B testing is ready you should get something like this:

  "flags": {
    "test-flag": true, // not an A/B testing
    "test-test-feature_name_variant": true
  "variants": {
    "test-feature_name_variant": {
      "name": "enabled" // or "disabled"

If instead your A/B testing is inactive you should get something like this:

  "flags": {
    "test-flag": true // not an A/B testing
  "variants": {}

Including extra data

Some times you need to include additional data to know how to handle each variant, this is useful because you can develop the variant implementation detached from the end value (that can be content that is not ready yet). You will then be able to develop and deploy your A/B testing and, other teams in charge of producing the content, will be able to test it without any other interaction with your code.

Extra data

Segment: measuring your new A/B testing

To measure if any of the variants of an A/B testing is significant we need to know three things:

To know the total number of users we send an event feature_flags each time a user gets a new set of feature flags, but to know which variant each user gets we need to format our feature flags as the following array:


For example, the previous feature flags will be format as:

  "test-flag", // not an A/B testing
  // or "test-feature_name_variant:disabled",

And send it in the prop featureFlags this way we can easily search feature flags and/or variants.

analytics.track(`feature_flags`, {
  featureFlags: [
    "test-flag", // not an A/B testing

Finally, you need to define your success metric (and time frame). This way the data team (or any other team) will be able to generate a dashboard to measure the success (or failure) of your test.


Copyright and related rights waived via CC0-1.0. Living