> ## Documentation Index > Fetch the complete documentation index at: https://docs.springwinter.dev/llms.txt > Use this file to discover all available pages before exploring further. # Amazon S3 Latency and Throughput Explained > Understand S3 latency, throughput, prefixes, parallel transfers, multipart operations, byte ranges, acceleration, caching, and measurement. Amazon S3 performance is a path, not one latency number. Client location, network capacity, concurrency, object size, key distribution, retries, and caching all affect the result. *Updated October 9, 2026.* **Measure the operation users experience, then tune the stage that limits it.** ## Start with the request path ```text theme={null} application work → DNS and connection setup → client network → S3 request processing → object bytes transferred → retries and validation → application consumes result ``` Time to first byte and total transfer time answer different questions. A metadata request is usually latency-sensitive; a 50 GB download is throughput-sensitive. A p50 alone hides slow requests. ## Prefixes create scaling dimensions S3 uses a flat key namespace. AWS documents at least 3,500 write-class or 5,500 read-class requests per second per partitioned prefix, with no documented limit on prefix count. These figures are request-rate guidance, not latency guarantees. Scaling is gradual. Rapid increases can return `503 Slow Down`, so clients should retry with backoff. Do not add random prefixes reflexively. Add distribution when measurements show a hot prefix or the required rate exceeds one prefix's guidance. | Technique | Best fit | Does not solve | | - | - | - | | Multiple prefixes | Very high request rates | Slow client code | | Parallel connections | Aggregate transfers | Tail-latency handling | | Multipart upload | Large uploads | Small-object overhead | | Byte-range GETs | Large or partial reads | Popular-object caching | | Transfer Acceleration | Long geographic paths | Local bottlenecks | | CloudFront | Repeated distributed reads | One-off writes | **Multiple prefixes increase aggregate request capacity only when the client distributes traffic across them.** ## Parallelism turns capacity into throughput A single connection can be constrained by round-trip time, congestion control, buffers, or runtime behavior. Parallel requests let a capable host use more network capacity. Increase concurrency gradually. Record throughput, latency percentiles, CPU, memory, connections, retransmissions, errors, and retries. Stop when throughput plateaus or tail latency becomes unacceptable. Reuse connections through the current AWS SDK. Configure bounded retries with exponential backoff and jitter. Immediate full-concurrency retries can amplify a transient overload. ## Multipart upload isolates work Multipart upload sends parts independently. Parts can run in parallel, and a failed part can retry without restarting the complete object. After all parts arrive, the client completes the upload and S3 publishes the object. Part size and concurrency interact. Very small parts increase request overhead; very large parts reduce parallelism and expand retry work. Incomplete parts remain billable until completed or aborted. **Multipart upload improves parallelism and retry scope; it does not guarantee lower latency for small objects.** ## Byte ranges accelerate reads The HTTP `Range` header returns part of an object. Workers can request non-overlapping ranges concurrently and combine them in order. Smaller ranges reduce repeated data after failure. For multipart-created objects, AWS recommends aligning ranges with original part boundaries when practical. Range reads also help formats whose indexes let applications fetch only needed sections. Parallel ranges are not always faster. Small objects, limited bandwidth, validation CPU, or too many requests can erase the benefit. ## Acceleration and caching solve different paths S3 Transfer Acceleration uses AWS edge locations and an optimized network path between an edge and the bucket. It targets long-distance transfers and adds charges. CloudFront caches eligible responses near viewers so repeated reads can avoid the S3 origin. Cache keys, TTLs, invalidation, object versioning, and origin configuration determine whether caching helps. Use Transfer Acceleration for long-distance movement to or from a bucket. Use CloudFront when distributed viewers repeatedly read cacheable content. **CloudFront reduces repeated origin reads; Transfer Acceleration optimizes the path to the bucket.** ## Measure the distribution Test representative Regions, networks, object sizes, operations, encryption modes, and concurrency levels. Capture DNS, connection, TLS, time to first byte, total duration, p50 through p99, throughput, status codes, retries, and client resource use. Warm and cold tests answer different questions. Connection reuse, DNS caches, CloudFront caches, and buffers can make later requests faster. Label conditions and change one parameter at a time. ## Frequently asked questions Not for ordinary workloads. S3 supports high rates with sequential names. Add well-distributed prefixes when measured traffic approaches per-prefix guidance or a hot prefix appears. Preserve useful listing, policy, and lifecycle boundaries. Use multipart upload for large objects when parallel transfer, resumability, or smaller retry units matter. Select part size and concurrency from measurements. Always complete or abort uploads, and consider lifecycle cleanup for abandoned parts. Choose by path. CloudFront serves repeatedly requested cacheable objects near viewers. Transfer Acceleration improves long-distance movement between clients and an S3 bucket through edge locations. They can coexist and should be compared using end-to-end measurements. ## Sources and further reading * [Optimizing S3 performance](https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance.html) * [S3 performance guidelines](https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance-guidelines.html) * [S3 performance design patterns](https://docs.aws.amazon.com/AmazonS3/latest/userguide/optimizing-performance-design-patterns.html) * [Multipart upload overview](https://docs.aws.amazon.com/AmazonS3/latest/userguide/mpuoverview.html) * [S3 Transfer Acceleration](https://docs.aws.amazon.com/AmazonS3/latest/userguide/transfer-acceleration.html) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.