Engineering
How sticky percentage rollouts work without storing users
MurmurHash3, a per-flag salt and basis points: why the same user lands in the same bucket across every SDK.
Draft: the full text of this post is not written yet
Percentage rollouts hash a stable attribute (the user key by default) together with a per-flag salt using MurmurHash3, take the result modulo 10,000 and compare it with the rollout expressed in basis points. Nothing about the user is stored: the same inputs always produce the same bucket, in every SDK.
The targeting concepts page has a calculator you can run.