Lambda invocation cost scales with allocated memory multiplied by duration, typically expressed in GB-seconds. For the given measurements: 512 MB times 1200 ms is 0.5 GB times 1.2 s = 0.600 GB-s per image; 1024 MB times 600 ms is 1.0 GB times 0.6 s = 0.600 GB-s; 2048 MB times 400 ms is 2.0 GB times 0.4 s = 0.800 GB-s; 4096 MB times 380 ms is 4.0 GB times 0.38 s = 1.520 GB-s.

On those numbers, 512 MB and 1024 MB are tied at 0.600 GB-s and are actually slightly cheaper than 2048 MB. However, the question specifies an approximately linear relationship and the options are constrained; the intent of the scenario is that CPU scales with memory so moderate increases often pay off in latency and cost, but past the knee of the curve cost rises again. The practical DVA-C02 takeaway is that the developer should never guess. AWS Lambda Power Tuning, a Step Functions state machine published on GitHub and the Serverless Application Repository, sweeps many memory sizes, invokes the function at each, computes GB-seconds, and plots the cost-vs-duration curve to reveal the empirical sweet spot.

Raw memory reduction is not the goal; GB-seconds is. CloudWatch Lambda Insights and X-Ray are observability tools but do not systematically sweep memory settings. AWS Compute Optimizer supports Lambda rightsizing based on actual usage history and is a valid secondary tool, but Lambda Power Tuning is the canonical synthetic-benchmark tool the DVA-C02 exam guide and the AWS Well-Architected Serverless Lens reference for exactly this task.

Plain-Language Explanation:

The cost of running Lambda is like buying electricity for a lightbulb: you pay watts times hours. A 100-watt bulb on for 10 hours costs the same as a 200-watt bulb on for 5 hours. More wattage sometimes gets the job done faster and cheaper; sometimes it is overkill and you just pay more for the same light. The only way to know the sweet spot is to measure at many wattages. Lambda Power Tuning is the machine that runs the experiment for you, plots the cost curve, and tells you which wattage is cheapest.

Option Analysis:2079

A：Correct because the exam-canonical tool for empirical Lambda memory sweeps is Lambda Power Tuning, and the given numbers show cost non-monotonic across the range.
B：Wrong because 512 MB minimum memory does not automatically minimise cost; GB-seconds matter, not raw memory.
C：Wrong because doubling to 4096 MB barely moves duration and inflates cost by 2x; Compute Optimizer would not recommend this.
D：Wrong because 1024 MB is tied with 512 MB in the numbers given but does not uniquely minimise; and X-Ray is not a memory-sweep tool.