MediLex operates a SageMaker real-time endpoint (ml.m5.large instances) for an NLP inference API. CloudWatch metrics reported an average of 800 invocations per minute over the last peak 5-minute interval and the metric InvocationsPerInstance is 100 invocations/minute. The endpoint currently runs 8 instances. The SRE team wants to right-size to meet the same peak traffic while maintaining the same per-instance load (target 150 invocations per instance). Using the CloudWatch InvocationsPerInstance or aggregate invocation data, how many instances should they provision to reach the 150 invocations/instance target?

The current peak load is 800 invocations per minute. To calculate the number of instances needed to meet this load at a target of 150 invocations per instance, divide the total invocations by the target invocations per instance: 800 invocations/minute / 150 invocations/instance = 5.33 instances. Since you cannot have a fraction of an instance, round up to the nearest whole number, which is 6 instances. 4 instances: This would result in 800/4 = 200 invocations/instance, exceeding the target of 150. 8 instances: This is the current number of instances, which results in 800/8 = 100 invocations/instance, below the target of 150. 10 instances: This would result in 800/10 = 80 invocations/instance, significantly below the target of 150 and over-provisioning.