No results for ""
EXPAND ALL
  • Home
  • API docs

Understanding sample ratios

Read time: 5 minutes
Last edited: Sep 18, 2024

Overview

This topic explains what sample ratios are and how to detect a sample ratio mismatch (SRM).

In your experiments, the sample ratio is the ratio between end users receiving each of your experimentation treatments. An SRM is a mismatch between the proportion of users expected in your experimentation treatments and the actual number of users in each treatment.

SRMs can occur if you are using JavaScript-based client-side SDKs. If LaunchDarkly alerts you of an SRM in one of your experiments, there are actions you can take to correct it.

Here's what an SRM alert looks like on an experiment's Results tab:

An SRM alert on an experiment's "Results" tab.
An SRM alert on an experiment's "Results" tab.

How to fix SRMs in JavaScript-based SDKs

SRMs are most commonly caused by outdated SDKs, event deduplication, and problematic application behavior.

Update your SDKs

Outdated SDKs are the most common cause of SRMs. If you observe an SRM and are using any of the LaunchDarkly JavaScript-based client-side SDKs, you may need to update your SDK.

If you are using JavaScript or React Web, confirm that you are using at least these versions:

  • Version 2.23.0 or newer of the JavaScript SDK
  • Version 2.28.0 or newer of the React Web SDK

These versions send the correct versions of events that LaunchDarkly needs for experimentation tracking. To learn more, read Experimentation and metric events.

Newer versions of LaunchDarkly SDKs replace users with contexts

A context is a generalized way of referring to the people, services, machines, or other resources that encounter feature flags in your product. To learn more, read Contexts.

Creating contexts and evaluating flags based on contexts is supported in the latest major versions of most of our SDKs. If you upgrade your JavaScript SDK or React Web SDK to version 3, those versions replace users with contexts. These new major versions also include the updates in versions 2.23.0 and 2.28.0, respectively, that address the most common causes for SRMs.

Event deduplication

Older versions of the JavaScript SDK, prior to version 2.23.0, deduplicate events by default. This can cause an SRM in the case where you evaluate the same flag before and after a call to identify.

Expand for an example of event deduplication with identify

If you add attributes to your user object after initialization and then call identify, event deduplication can cause a non-random loss of experimentation events.

Here's how this can happen:

  1. Your code initializes the JavaScript SDK.
  2. An end user encounters a flag in your application. The user does not match your experiment rule, and LaunchDarkly serves a flag variation to the user.
  3. Your application learns more about the end user. Your code adds attributes to the user object, without changing the user key, and calls identify.
  4. Because of the identify call, LaunchDarkly re-evaluates the flag for the same user. Based on the additional attributes, the user now matches your experiment rule and is eligible for the experiment.
  5. The user is randomly assigned to the same flag variation that they were previously served.

If these steps happen within the SDK's deduplication window of five minutes, the user will not be counted as part of the experiment. If, in step 5, the user is randomly assigned to a different flag variation than they were previously served, then they will be counted correctly as part of the experiment. Because some users are being counted and some users are not, you will have an SRM.

Older versions of the React Web SDK, prior to version 2.28.0, have a similar issue. Furthermore, in the React Web SDK LDProvider calls allFlags after initializing the client. This can potentially exacerbate issues with event deduplication.

To fix this, we strongly recommend upgrading your SDK version. If you can't upgrade your SDK version, you can set the allowFrequentDuplicateEvents option to true.

Expand for allowFrequentDuplicateEvents code sample

Here's how to set the allowFrequentDuplicateEvents option in your JavaScript SDK to true:

const options = { allowFrequentDuplicateEvents: true };
const client = LDClient.initialize('client-side-id-123abc', user, options);

The allowFrequentDuplicateEvents option has been deprecated in the JavaScript SDK version 2.23.0 and React Web SDK version 2.28.0. It has been removed in version 3 of the JavaScript and React Web SDKs.

Event loss due to application behavior

Some experimentation implementations can lose events disproportionally for JavaScript-based client-side SDKs. For example, if your experiment is an A/B test where the "A" does nothing and the "B" immediately redirects your end users to another page, it's possible that more events from the "B" variation will be lost.

The LaunchDarkly SDKs send events immediately when the page unloads. However, some browsers, such as Safari on iOS, do not respect asynchronous calls on unload. If LaunchDarkly cannot record all of the events for your experiment, then your experiment will have an SRM.

Check to see if one of your experimentation treatments results in users leaving the page more quickly than the other treatments. You may need to adjust your experimentation treatments to account for this behavior.

How SRMs occur

There are many ways an SRM can occur, but the most common is that the randomization method is incorrect.

Imagine you have an A/B test with 50% of users assigned to treatment A and 50% assigned to treatment B, you should expect that with a sufficient sample size, the split is roughly 50/50. If you have 10,000 users in your experiment iteration, you should expect around 5,000 users in treatment A and 5,000 in treatment B. Because the sampling is random, this split will not be exact. However, it should be close. You should not have 4,000 users in treatment A and 6,000 users in treatment B.

LaunchDarkly uses statistics to determine if the sample ratio is more different than what you should expect from a random sampling. If there's a 99% or greater probability that the units, or users, in the experiment are not arriving at the same rate as expected given how you've allocated the traffic, then LaunchDarkly displays a warning noting that your experiment results are almost certainly invalid.