Question 23
Domain 4: Model DeploymentWhy might a team start with a 90/10 traffic split instead of switching 100% of traffic to a new model immediately?
Correct answer: A
Explanation
A 90/10 split is a staged rollout that “lowers rollout risk” because only a small portion of users see the new model at first. It also lets the team “observe the candidate under real production traffic,” which matches safe deployment and model evaluation practices before shifting all traffic.
Why each option is right or wrong
A. It lowers rollout risk while letting the team observe the candidate under real production traffic
A 90/10 split is used as a controlled canary-style deployment: only 10% of requests are exposed to the new model, so any regression is contained while 90% of traffic continues on the stable path. In practice, this lets the team measure live latency, error rates, and output quality under real production conditions before increasing exposure to 100%.
B. It eliminates the need for monitoring altogether
Monitoring is still essential to compare quality, latency, errors, and drift during rollout.
C. It guarantees the new model will become champion automatically
Traffic splitting tests a candidate; it does not automatically promote it to champion.
D. It prevents the old model from serving any requests
A 90/10 split means the old model still serves most requests, not zero.