MediaWatch runs a real-time video classification model and wants to alert when model execution time rises. Which CloudWatch metric should be used to detect increased model container inference time only, and what is the recommended way to alarm on sustained increases for a particular variant of an endpoint?

The correct answer identifies 'ModelLatency' as the metric that specifically measures the time spent inside the model container, which is crucial for detecting increases in model inference time. Using 'EndpointName' and 'VariantName' dimensions ensures the alarm is specific to a particular model variant. Alarming on a p90 (or other percentile) statistic over a 5-minute period with multiple evaluation periods (e.g., 3) effectively detects sustained increases, preventing false alarms from transient spikes. 'OverheadLatency' measures the time spent outside the model container, such as network overhead, not the model's execution. 'InvocationDuration' measures the total time from request to response, including network and SageMaker overhead, not just model compute. Alarming on a global 'ModelLatency' without dimensions would not isolate the issue to a specific variant, and alarming on a 'Minimum' dropping would indicate decreased latency, not increased.