Skip to main content

FAQ

Below are some frequently asked questions about GrowthBook.

Do users always get assigned the same experiment variation?

GrowthBook SDKs use deterministic hashing to ensure the same user always gets assigned the same variation in an experiment.

In a nutshell, GrowthBook hashes together the hashAttribute (the user attribute used to assign a variation, ex: user id) and the experiment trackingKey which produces a decimal between 0 and 1. Each variation is assigned a range of values (e.g. 0 to 0.5 and 0.5 to 1.0) and the user is assigned to whichever one their hash falls into.

This does mean, if you change the experiment configuration, some users may switch their assigned variation. For example, if someone has the hash 0.49 and you adjust the weights to a 40/60 experiment, the variation ranges become 0 to 0.4 and 0.4 to 1.0. In this case, the user was previously in the control group, but will now be in the variation.

GrowthBook will detect issues like this and will remove users who see both variations from the analysis automatically. However, to keep things simple and safe, we recommend not relying on this and treating experiments as immutable once they are running.

It's important to note that the above only applies when changing the traffic split between variations. If you keep the split the same, but increase the percent of traffic included, users will not switch variations. For example, if you are running a 50/50 experiment on 20% of traffic, the variation ranges will be 0 to 0.1 and 0.5 to 0.6. Users outside those ranges will be excluded from the experiment. If you increase the percent of traffic to 40%, but keep the 50/50 split, the ranges will become: 0 to 0.2 and 0.5 to 0.7. As you can see, no users switch variations. Instead, some users who were previously excluded are now part of the experiment.

What do I use for an "id" attribute in the SDK if my users aren't logged in?

If your application has both logged-in and anonymous users, we recommend using two identifier attributes:

  • id which is the database identifier of logged-in users (or empty string for anonymous)
  • deviceId (or sessionId, etc.) which is a random anonymous hash, persisted in a cookie or local storage. This should always be set for both anonymous and logged-in users.

If your application only has anonymous users (e.g. a static marketing site), then we recommend a single id attribute which, similar to deviceId or sessionId above, is a random hash persisted in a cookie or local storage.

Why is the trackingCallback not firing in the SDK?

The trackingCallback only fires when a user is included in an experiment. If you're expecting to be included and you're still not seeing the callback fire, it's likely for one of the following reasons:

  • You are missing the hashAttribute for the experiment. For example, when you are splitting users by "company", but the company attribute is empty.
  • The feature is disabled for the environment you are in (dev/prod)
  • The experiment has reduced coverage. For example, if it's only running for 10% of users and you are in the 90% that are excluded.
  • There is another feature rule that is taking precedence over the experiment.

If you are using the Javascript or React SDK in a browser environment, you can install the GrowthBook DevTools Chrome Extension to help you debug.

Note: To use the plugin, you will need to pass enableDevMode: true when creating your GrowthBook instance.

const growthbook = new GrowthBook({
enableDevMode: true,
})

How do I run an A/B test in GrowthBook?

The recommended way to run an A/B test is by using Feature Flags and our SDKs.

  1. Create a feature in GrowthBook (e.g. new-signup-form) with an A/B Experiment rule
  2. Use our SDKs to serve the different variations
    if (growthbook.feature("new-signup-form").on) {
    // Variation
    } else {
    // Control
    }

What is the best way to redirect users to a URL based on their experiment variation?

Experimenters often want to display different URLs to different users. Unfortunately, waiting to look up a user's experiment variation and rendering a page accordingly can lead to undesirable effects like flickering in the user interface. If only one of the variations suffers from flickering or a delay in loading, these problems can also lead to bias.

In general, we suggest three approaches, ranked from the least risky to the most risky in terms of user experience and potential bias:

  1. Render page A or page B conditionally using the GrowthBook SDK within the backend controller logic or middleware layer. This method is quick, unbiased, and avoids any flickering or sudden redirection.
  2. Render page A or page B conditionally using the GrowthBook JS SDK on the front-end. This option is nearly as good as the first one, and while you might be able to implement this without any flickering for users, it might have a slight loading delay. It can be unbiased if both pages A and B result in the same loading delay.
  3. Perform front-end redirection using JS code snippets in the visual editor with the command window.location.href="mytarget.html". This approach is problematic but can be acceptable if executed carefully. It introduces noticeable flickering before and after redirection. For fairness, it's best to initiate the redirection from a neutral or minimally loaded page for both variations A and B. This helps avoid introducing bias—for instance, preventing a scenario where page A appears more favorable simply because it loaded faster due to the implementation of the experiment.

How much traffic do I need to run A/B tests?

What matters most for A/B testing is not traffic, but conversions. The general rule of thumb is to have at least 100-200 conversions per variation before you might start reaching significance.

So that means if you do 50 orders per week and that's the metric you are trying to optimize, you'll need to run a simple 2-way A/B test for at least 4-8 weeks. If you run a 3-way test, it will take 6-12 weeks.

Can I run multiple A/B tests at a time?

Yes! In fact, we recommend running many experiments in parallel in your application. Most A/B tests fail, so the more shots-on-goal you take, the more likely you are to get a winner. Running tests in parallel is a great way to increase your velocity.

Now it's possible your experiments might have interaction effects, but these are actually pretty rare in practice. One example is if one test is changing the text color on a page and another test is changing the background color. Some users might see end up seeing black text on a black background, which is obviously not ideal. For these rare cases, you can use Namespaces to run mutually exclusive experiments.

As long as you apply a little common sense to avoid situations like the above, running multiple experiments has low risk and really high reward.

What is the difference between a Dimension and a Segment?

A dimension is a user attribute that can have multiple values. Some examples are country, account_type, and browser.

A segment is a specific group of users. Some examples are visitors in the US, premium users, and chrome users.

Dimensions are used to explore experiment results. For example, you can use a country dimension to see which countries had the highest conversion rates. Or an account_type dimension to see if there was a significant difference in how free vs paid users behaved. Or a browser dimension to detect any browser-specific bugs in your implementation.

Segments can apply a filter to results, usually to compensate for bad data. For example, if your experiment was only visible to premium users, but your database inaccurately shows that free users were also included, you could apply a premium users segment to only include those who were actually exposed to the test. Ideally, you could just fix the underlying data, but that's often not feasible so segments provide a quick and dirty alternative.

Which docker image tag should I use when self-hosting?

We recommend using the latest tag for both dev and production self-hosted deployments. This tag represents the latest stable build of GrowthBook and is what the Cloud app uses.

Specific version tags (e.g. v1.1.0) are only released periodically (about once a month) and you will miss out on the many bug fixes and features added between major releases.

We also recommend updating the image regularly. You can do that by downloading the latest image (docker pull growthbook/growthbook:latest) and restarting the container.

What are the hardware requirements for self-hosting GrowthBook?

The GrowthBook application is very lightweight and efficient. For most usecases, 2GB of memory is sufficient even for large production deployments.

GrowthBook only deals with aggregate data and the bulk of the processing is offloaded to your data source. Because of this, you can easily analyze terrabytes of data from your laptop or a small container in the cloud.

If you are using feature flags, we strongly recommend adding a caching layer between the GrowthBook API and your application in production. This will also help you stay within the limits of our Fair Use Policy. Some of our code examples implement caching. We offer a pre-built GrowthBook Proxy server you can run that handles caching and invalidation automatically. You can also setup your own custom system using a CDN or distributed cache like Redis.

I can't upload to S3/getting 400 error when uploading to S3?

  • Make sure you've correctly set the S3_BUCKET and S3_REGION environment variables
  • Enable bucket ACL and set ownership to Bucket owner preferred: read more here.
  • Make sure the S3 bucket is publically accessible
  • Make sure CORS settings are correct. Add your URLs to the AllowedOrigins array or set to "*"

How do I use the Chrome DevTools Extension?

  • Make sure you are using the React or Javascript SDK
  • Install the Chrome DevTools Extension
  • Pass the enableDevMode: true option into the GrowthBook constructor
const growthbook = new GrowthBook({
enableDevMode: true,
})

My old exported notebook stopped working. How can I fix it?

There's a good chance that the SQL we are exporting and your version of our Python stats library, gbstats, are out of sync. In February of 2023 we updated our SQL engines and gbstats library to only use sums and sums of squares, rather than averages and standard deviations.

If the queries in your notebook return averages and standard deviations (using AVG and VAR SQL operators as part of the __stats CTE), then you need to run that notebook with gbstats version 0.3.1. You can download this from PyPI here using pip install gbstats==0.3.1 and ensure that your kernel uses that version of gbstats. Ideally in this case you can redownload the notebook and use the new gbstats library (0.4.0 or newer). You can download a new notebook by navigating to your experiment in GrowthBook and clicking Download Notebook again. This should now use the updated SQL and gbstats syntax. Then, if you install gbstats 0.4.0 or later, everything should work as expected.

If the queries in your notebook return sums and sums of squares as part of the __stats CTE but your notebook is still erroring, then you probably have an old gbstats version installed and need to update to 0.4.0 or later. You can download this from PyPI here using pip install gbstats and ensure that your kernel uses that version of gbstats.

My features aren't refreshing as expected. What can I do?

Our SDK's implement a stale-while-revalidate approach to cacheing with a configurable time-to-live (TTL) value.

This means that if the feature payload is considered stale (i.e. more than the TTL amount of time has passed since it's been updated), the next request will return the stale features and refetch an update asynchronously so that on the next request, the features will be up to date. You can learn more about how our SDK's implement this in detail here.

If you would like something more real-time than this stale-while-revalidate approach, you may want to consider implementing the GrowthBook Proxy on your self-hosted instance.

How do I configure environments in the SDK?

When you create an SDK connection, it is linked to a specific environment. You can learn more about environments here.

How do I disable the on-screen celebrations?

Throughout the GrowthBook application, we randomly celebrate key milestones like launching experiments with on-screen confetti. If you'd like to disable this, you can click on your avatar in the top right corner and select "Edit Profile". From there, you can disable the toggle for "Allow Celebrations". Please note this is persisted in your browser's local storage, so if you clear your browser's local storage, you will need to disable this again.


Can't find your question?

If you can't find an answer to your question above, please let us know so we can help you out and improve the docs for future users!

You can join our Slack channel for the fastest response times.

Or send an email to hello@growthbook.io if Slack isn't your thing.