เลื่อนโมเดลที่ชนะหรือย้อนกลับ

Your A/B test has run its full window and the numbers are in. Now comes the decision: promote the challenger or roll it back to the champion. 🚦

Pull up the success criteria you wrote before the test. The challenger wins only if it beats the threshold on the primary metric and breaks no guardrail.

To promote a model is to make it the new champion. With a registry, that often means moving the winning version into the Production stage.

MLflow lets you flip the winning version to Production in one call, so your serving layer picks it up.

Even a clear winner deserves a careful rollout. Raise its traffic from ten percent toward one hundred gradually while you watch the live metrics.

Roll back if the challenger loses, ties, or trips a guardrail like a latency spike. A tie is not a win, so the safe default is to keep the champion.

A clean rollback is just pointing traffic back at the old version. Because the champion is still registered, recovery should take seconds, not a rebuild. ⏪

Never delete the previous champion the moment you promote. Move it to the Archived stage instead, so a fast revert is always one step away.

Promotion is not the finish line. Keep monitoring the new champion, since live behavior at full traffic can still surprise you after the test ends.

Log who decided, the metrics, and the verdict. This audit trail lets you explain later exactly why this model went live or got reverted.

Rolling back is not failure. You protected users from a worse model and learned something real, which is exactly what the experiment was for. ✅

Let us decide what to do after the results land.

Decide against your written plan: promote a clear, guardrail-safe winner via the registry, ramp gradually, and keep the old champion for an instant rollback. 🎯