MLA-C01 Practice Question 87: Scheduled Endpoint Scaling
AWS Certified Machine Learning Engineer – Associate (MLA-C01) · Test 9 · question 7 of 10
A payroll company's tax-classification endpoint uses a target tracking policy on InvocationsPerInstance. Every weekday at 09:00, when client offices open, traffic jumps about tenfold within a minute. New instances take several minutes to launch and pass health checks, so for the first 10 minutes of each weekday latency spikes and some requests time out. Traffic is steady the rest of the day and near zero overnight and on weekends.
What should the engineer add to prevent the morning latency spikes?
Show answer and explanation
Answer:
Target tracking and step scaling are *reactive*: they add capacity after a metric rises, and new instances still take minutes to be ready. This spike comes at a known time every weekday. SageMaker AI endpoints support scheduled scaling through Application Auto Scaling. Scheduled actions can run on a recurring schedule, and after one runs, the scaling policy keeps adjusting capacity as the workload changes. Raising the minimum capacity before 09:00 means the instances are already serving when the traffic arrives, and target tracking still handles the rest of the day (option C).
- A is wrong because a lower target still waits for traffic to rise before launching instances, so the 09:00 jump still arrives before new capacity is ready. It also over-provisions all day. - B is wrong because step scaling is reactive too. Adding many instances at once does not make them launch faster. - D is wrong because the cooldown governs how soon the policy can scale out *again*. The delay here is how long the first instances take to start, which the cooldown does not change.
Reference: Auto scaling policy overview - Schedule-based scaling