問題文
A service handling 300,000 requests a day must adopt a new model version. The requirements are: no regression reaches more than a small share of users, the decision is based on measurement, and the previous behavior can be restored within minutes. Which plan meets all three?
選択肢
- Deploy the new version to all instances after it passes the offline evaluation, and keep the previous image available for redeployment; an offline evaluation that covers the release criteria establishes the absence of regressions before any user is exposed, and redeployment then covers the remaining risk.
- Route by user identifier so that a fixed set of users always gets the new version.
- Keep the previous version loaded and routable, send an increasing share to the new one while comparing fixed metrics at each share, and switch the routing back if a metric crosses its threshold.
- Run the new version on copies of production requests without returning its output, then switch all traffic.