Throughput is the rate at which a service processes operations over time. In AWS storage and caching systems, it commonly describes reads and writes per second, while network throughput describes bytes received or sent. section 7.8 section 17.3.1
S3 assigns default limits of 3,500 writes and 5,500 reads per second per partitioned prefix. Because object keys use prefixes rather than real directories, naming affects the throughput available to a workload. section 7.8
Increase or manage throughput by matching the architecture to the bottleneck:
- Distribute S3 objects across additional prefixes when one prefix cannot handle the required read or write rate. section 7.8
- Use sharding when cached data exceeds one node's memory or when the system needs more throughput. section 11.2
- Add Redis shards to divide requests across nodes; with even distribution, two shards handle roughly half the requests each. section 11.2.4
- Use a proportional CloudWatch metric, such as network bytes, for target tracking when scaling EC2 instances. section 17.3.1
Splitting an S3 workload
A workload using only archive/image1.png through archive/image4.png shares one archive prefix and its limit. Moving objects under archive/2021 and archive/2022 distributes requests across prefixes and can double the available read throughput. section 7.8
archive/2021/image1.png archive/2021/image2.png archive/2022/image3.png archive/2022/image4.png
Compare a single archive prefix with year-based prefixes to see how object-key structure distributes S3 requests. section 7.8
