AWS has introduced S3 Files, a service that mounts an S3 bucket as a native, high-performance file system on compute instances. For data engineers, this changes a fundamental architectural constraint: you no longer need to choose between S3's scale and a file system's interactive, mutable interface.

The system presents S3 objects as files and directories, supporting full NFS v4.1+ operations. Writes on the mount are synced back to S3 as new object versions. The architecture uses an EFS-based cache for active data, delivering ~1ms latency, while serving cold or large sequential reads directly from S3 to optimize throughput and cost. Engineers can configure pre-fetching and control what data resides in the cache.

From a design perspective, this service targets workloads where multiple compute resources—like distributed ML training jobs or containerized applications—require shared, low-latency access to the same S3 dataset without manual copying. It uses IAM for access control and POSIX permissions mapped to object metadata.

The trade-off is operational complexity versus data unification. While it simplifies architectures by making S3 a direct source for file-based tools, you now manage mount targets and cache behavior. It's not a direct replacement for high-performance file systems like FSx for Lustre, but rather a bridge for workloads already tied to S3's economics and durability.

Pricing is based on cache storage, file system operations, and S3 request charges for synchronization. The service is available now across all commercial AWS regions.

Source: AWS