A company hosts a web application on Amazon EC2 instances that are part of an...

For short, frequent daily spikes (15–20 minutes), the most cost-effective scaling behavior typically comes from target tracking scaling driven by a request-based metric, because it continuously adjusts capacity to maintain a chosen target value rather than adding/removing capacity in coarse steps. Option D uses a custom CloudWatch metric representing the number of GET requests (sourced from ALB access logs, application instrumentation, or request counting logic) and applies target tracking so the Auto Scaling group scales out and in proportionally to demand. This helps avoid both under-provisioning (bad performance during spikes) and over-provisioning (wasted cost between spikes), which is exactly what cost-optimized elasticity aims for.

Option A (simple scaling) is usually less cost-efficient for bursty traffic because it relies on fixed adjustments and alarm thresholds, often with cooldowns. With 15–20 minute spikes, simple scaling can react too slowly, overshoot, or oscillate, leaving extra instances running after the spike or failing to ramp fast enough at the start. Option B (predictive scaling) is intended for regular, forecastable demand patterns, but it’s not always the most cost-effective choice for repeated short spikes because it can pre-scale conservatively and may require more historical data and tuning; it is often paired with dynamic scaling anyway. Option C uses UnhealthyHostCount, which is a health signal—not a demand signal—and scaling on unhealthy hosts can increase cost without addressing request volume; it may also mask underlying application issues rather than scaling to meet load.

Because the requirement explicitly allows standard or custom metrics and focuses on scaling based on request volume, D best matches a cost-efficient, responsive approach that directly tracks incoming workload.