Use On-Demand warm pools for critical data and a separate Spot group scaling on a CloudWatch agent memory metric

Why the Answer Is Correct

The two data types have opposite requirements, so they need separate scaling configurations. The critical data cannot tolerate interruption and must scale out quickly on real-time demand, so it stays in the existing Auto Scaling group on On-Demand Instances and gains a warm pool whose stopped instances are pre-initialized, which is the feature that actually reduces scale-out latency; a new launch template version with detailed monitoring enabled supports the visibility needed, and On-Demand rather than Spot is required because interruption is not acceptable (B). The noncritical data can withstand interruption and is driven by memory consumption, so it moves to a second Auto Scaling group on Spot Instances to gain the cost benefit; its launch template installs the unified CloudWatch agent and configures a custom memory utilization metric, which supplies the signal a target tracking scaling policy needs (D). Splitting the two tiers this way satisfies both the availability and the cost requirement without compromising either. B and D are the correct combination.

Why the Other Options Are Wrong

A modifies the existing Auto Scaling group with a warm pool and detailed monitoring but uses Spot Instances for the critical data. The requirement states that the critical data analysis cannot tolerate interruption, and Spot Instances carry exactly that risk; jamesf and trungtd both identified this as disqualifying, and a warm pool of Spot capacity can be reclaimed. C modifies the existing group with a lifecycle hook that ensures bootstrap completes and the application is ready before instances are registered, and enables detailed monitoring in a new launch template version. A lifecycle hook improves deployment safety and readiness but does nothing to reduce scale-out latency, because new instances still have to be launched and initialized on demand; the warm pool in B is what reduces that latency. E creates a second Auto Scaling group and chooses the predefined memory utilization metric type for the target tracking policy while using Spot Instances. trungtd stated directly that AWS Auto Scaling does not provide a predefined memory utilization metric type, so the policy in E has no valid metric to track; this is exactly the gap option D closes by installing the CloudWatch agent and configuring a custom memory utilization metric. B and D are correct.

Community Comment Notes

Community voted B,D (93). trungtd gave the decisive technical fact that AWS Auto Scaling does not provide a predefined memory utilization metric type, which eliminates option E and explains why option D must install the CloudWatch agent and publish a custom metric. jamesf explained that option B creates a warm pool with On-Demand Instances for the critical data to address low-latency scaling and that option D uses Spot Instances for the noncritical data. TEC1 was the sole dissenter, favoring B and E, and argued that On-Demand is reliable while Spot carries termination risk; that reasoning supports B but does not overcome the absence of a predefined memory metric type in E. No alternative received substantive support.