Shadow, Canary, A/B with the Registry

Patterns that use two aliases

Once your serving layer can route on a header or a percentage, the registry alias mechanism becomes a deployment primitive:

• Shadow: alias @champion serves traffic, @shadow receives a copy without affecting the user. Compare offline.
• Canary: route 1% → 5% → 25% to @candidate, the rest to @champion. Roll back instantly by reverting the alias.
• A/B: random 50/50 split between @a and @b; collect outcome metrics and pick a winner statistically.

MLflow does not implement traffic routing — that's the serving platform's job (KServe InferenceGraph, Istio, SageMaker production variants). MLflow's role is to give you the stable, swappable identifiers the router consumes.

Aliases as deployment knobs are like theatre lighting cues: you don't move the stage, you change which lamp is on. Wrong cue? Reset the alias and the audience sees the original scene again in milliseconds.