Block, File, and Object Storage for Enterprise Analytics Pipelines: A Semantics-Aware AWS Performance Benchmark

Abstract

Enterprise analytics pipelines increasingly span block, file, and object storage, but benchmark comparisons often ignore the access semantics that determine how each service is actually used. We evaluated all three storage classes on AWS using six analytics workloads covering sequential scans, metadata-heavy transforms, checkpointing, random updates, and large-object exchange. Tests were repeated across three data scales and two concurrency levels. Object storage delivered the highest cost efficiency for large sequential reads but incurred substantial overhead for fine-grained update patterns. File storage provided the most consistent end-to-end runtime for workflows requiring shared namespace semantics, while block storage achieved the lowest p99 latency for transactional scratch operations. At high concurrency, metadata-bound file workloads degraded more sharply than throughput-oriented object workloads. No storage class was uniformly superior across the pipeline. A semantics-aware placement strategy that matched stages to their dominant access pattern reduced total workflow time by 17% and estimated storage cost by 13% relative to a single-service baseline. Storage selection for analytics should therefore be based on operation semantics and pipeline stage rather than headline throughput figures alone.