How randomization works in Amplitude Experiment

This article helps you:

Understand the process Amplitude Experiment uses to randomly assign users to experiment variants

Amplitude Experiment uses deterministic randomization of variations. This randomization uses the bucketing key selected in the UI, as well as the bucketing salt of the flag. In most cases, the Amplitude ID is used as the bucketing key.

This randomization model is two-dimensional: Amplitude Experiment performs a murmur3 hash of "bucketingSalt/amplitude_id" , and uses that hash to determine first, whether a specific user should be assigned a variant, and second, which variant that user should be assigned.

Note

In this article, / refers to the backslash character. It doesn't imply mathematical division.

For the first stage (the initial assignment), Experiment divides users into 100 buckets based on the value of mod(murmur3_x86_32("bucketingSalt/id", 100). If the bucket returned is less than the percentage rollout, then the user is bucketed into the experiment, in which case the user will be assigned a variant. If the user isn't bucketed into the experiment, they see the fallback variant instead.

For the second stage (variation assignment), Experiment will take all users bucketed into the experiment and assign each a variant according to the value of floor(murmur3_x86_32("bucketingSalt/id"), 100).

Variants are associated with values between 0 and 42949672, based on their weights. For example, in a two-variant experiment, if Variant A has weight 1 and Variant B has weight 1 (for example, each variant should receive 50 percent of the experiment’s traffic), a user would be assigned to Variant A if the value of the hash lies between 0 and 21474835. The user would be assigned to Variant B if that value lies between 21474836 and 42949672.