Canarying gets you a 1/n treatment group, but it might be skew geographically (all affected users are near the canary’s datacenter). You need a percentage in a feature flag if 1/n is too big and you want, e.g., 0.1% of traffic.