0Pricing
MLOps Academy · 课时

推动获胜模型上线或回滚

部署更好的模型，或干净利落地恢复原版本

推动获胜模型上线或回滚 是 CoddyKit 上的免费 MLOps Academy 课时。 这是第 4 节课，共 4 节。 你可以在下方免费阅读本课时的完整内容 — 然后在浏览器中使用内置代码编辑器和全天候 AI 导师进行实践。 这是 MLOps Academy 学习路径的一部分，你的进度在网页和 CoddyKit 应用中同步。 MLOps Academy 课程共包含 4 节课。

本课时的部分内容尚未翻译，以英文显示。

The Test Is Over

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. 🚦

Decide Against Your Plan

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.

Promotion Means Production

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.

Promote via the Registry

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

from mlflow import MlflowClient

client = MlflowClient()
client.transition_model_version_stage("fraud", version=7, stage="Production")

Ramp, Do Not Slam

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

When to Roll Back

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.

Rollback Must Be Instant

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. ⏪

Keep the Old Champion

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.

Watch After You Ship

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.

Record the Decision

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

A Losing Test Still Wins

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

Quick Check

Let us decide what to do after the results land.

Recap

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. 🎯

常见问题解答

「推动获胜模型上线或回滚」课时是免费的吗？

是的 — 「推动获胜模型上线或回滚」的完整文本可在网页上免费阅读。要进行交互式练习（内置代码编辑器和全天候 AI 导师）并解锁 MLOps Academy 课程的其余内容，请升级到 CoddyKit PRO。 MLOps Academy 课程共包含 4 节课。

「推动获胜模型上线或回滚」这节课中我会学到什么？

部署更好的模型，或干净利落地恢复原版本 你通过在浏览器中直接运行的动手代码来练习 MLOps Academy，全天候 AI 导师会在你学习这节课的过程中回答你的问题。

学习 MLOps Academy 需要有经验吗？

无需任何先前经验。CoddyKit 上的 MLOps Academy 课程适合初学者到高级学习者，你可以从这里开始或从头开始，按照自己的节奏学习。 这是第 4 节课，共 4 节。

「推动获胜模型上线或回滚」课时需要多长时间？

大多数 CoddyKit 课程大约需要 5–10 分钟。每节课都很精短且互动，所以你能稳步进步，并在网页和应用中从离开的地方继续。

我能在这节 MLOps Academy 课中编写并运行代码吗？

能。每节 MLOps Academy 课都包含内置代码编辑器，你可以在浏览器中直接编写并运行真实代码，并获得即时 AI 反馈 — 无需本地设置。

此课程中的所有课时

  1. 在不同模型版本之间分配流量
  2. 选择真正重要的指标
  3. 正确解读显著性
  4. 推动获胜模型上线或回滚
← 返回 MLOps Academy