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/n 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. The tracking issue is similar to a feature flag rollout issue, and is also used to track the status of an experiment. 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.

Implementing an experiment

There are currently two options when implementing an experiment.

One is built into GitLab directly and has been around for a while (this is called Exerimentation Module), and the other is provided by gitlab-experiment and is referred to as Gitlab::Experiment -- GLEX for short.

Both approaches use experiment feature flags, and there is currently no strong suggestion to use one over the other.

Feature Experimentation Module GLEX
Record user grouping Yes No
Uses feature flags Yes Yes
Multivariate (A/B/n) No Yes

Historical Context: Experimentation Module was built iteratively with the needs that appeared while implementing Growth sub-department experiments, while GLEX was built with the findings of the team and an easier to use API.