An application routinely processes a large number of Amazon S3 GET requests each second.

Amazon S3 automatically scales to handle high request rates, but request parallelism is optimized when objects are distributed across multiple prefixes. AWS documentation explains that S3 internally partitions data by key name, and distributing objects across multiple prefixes allows requests to be spread across multiple partitions.

When many GET requests target objects within the same prefix, they may contend for the same internal partition, limiting throughput. By designing object keys with multiple prefixes—such as including hashed or randomized components—applications can significantly increase parallel request handling.

Options A and D do not address S3 request partitioning. Global Accelerator improves network latency for supported endpoints but does not increase S3 request parallelism. Direct Connect improves network consistency but does not change how S3 scales requests internally. Option B worsens the problem by concentrating traffic into a single prefix.

Therefore, partitioning objects by prefixes is the correct and AWS-recommended solution.