Amazon S3 Files
A POSIX file-system interface over general-purpose S3 buckets, launched April 7, 2026. Any bucket can be mounted as an NFS v4.1 or v4.2 endpoint from EC2, ECS, EKS, or Lambda, giving workloads open/rename/edit-in-place semantics with ~1ms hot-file latency via an underlying Amazon EFS caching tier, while large sequential reads and byte-range requests pass through directly to S3.
Summary
A POSIX file-system interface over general-purpose S3 buckets, launched April 7, 2026. Any bucket can be mounted as an NFS v4.1 or v4.2 endpoint from EC2, ECS, EKS, or Lambda, giving workloads open/rename/edit-in-place semantics with ~1ms hot-file latency via an underlying Amazon EFS caching tier, while large sequential reads and byte-range requests pass through directly to S3.
S3 Files ends the twenty-year pattern of copying data from S3 into EBS or FSx for any workload that needs file semantics. The immediate beneficiaries are agentic AI systems (Claude Code, LangGraph orchestrators, multi-step model pipelines) which persist memory and share state using ordinary Python file I/O. Rather than architecting around the S3 REST API, agents treat the bucket as a mutable working directory.
- NFS close-to-open consistency vs S3 atomic-PUT strong consistency is bridged by 60-second write aggregation. Writes are visible via NFS immediately but don't land in the bucket via REST API until the next flush.
- Concurrent NFS + REST modifications resolve via a strict "S3 wins" policy — direct API writes silently overwrite in-flight NFS edits.
- Rename is not free. Under the hood it's copy + delete, so high-churn rename patterns incur real cost. Design around append/close rather than rename-as-commit.
- Metadata-only cache mode is the right call for large sequential workloads (pretraining video, uncompressed text corpora) — full file caching just wastes the EFS tier.
- Locking is advisory only. Mandatory locks aren't supported, and direct S3 REST writes bypass NFS locks entirely. Don't rely on the lock primitive for cross-tenant isolation.
- Glacier / Deep Archive objects don't appear under the mount until restored — surprising failure mode for migration scripts that walk a tiered bucket.
- Object key length still caps at 1,024 bytes, no relief from the file-system layer. Deeply nested paths with long names can hit the ceiling without warning.
- Max 128 mount targets per bucket, no exception path.
- Windows is not supported — FSx for Windows File Server is the AWS-blessed alternative.
- Object keys that aren't valid POSIX filenames silently don't appear under the mount, even though they're listable via the S3 API.
Pricing posture: EFS cache tier $0.30/GB-mo + reads $0.03/GB + writes $0.06/GB + metadata operations $0.005 per 1,000 + standard S3 rates underneath. The economics favor "small hot working set on top of a large cold bucket" — a 1 PB bucket with 1 TB active pays cache rates on the 1 TB only.
Companion change you'll see at the same time: SSE-C is disabled by default on new buckets after April 6, 2026 — AWS steering customers to KMS CMKs which carry rotation, audit, and recovery semantics SSE-C never had. Existing SSE-C buckets keep working.
implementsNFS v4.1 — the mount protocolimplementsS3 API — the underlying bucket protocolenablesAgent-Safe Views — agents persist memory with standard file I/OsolvesLack of Atomic Rename — surfaces rename at the NFS layersolvesCold Scan Latency — EFS cache fronts hot files at ~1msconstrained_byS3 Consistency Model Variance — close-to-open + atomic-PUT reconciliation
Definition
A POSIX file-system interface layered over general-purpose S3 buckets, launched April 7, 2026. Any existing bucket can be mounted as an **NFS v4.1 / v4.2** endpoint from EC2, ECS, EKS, or Lambda, giving compute workloads file semantics (open, rename, edit-in-place, lock) without rewriting code to the S3 REST API. Backed by an Amazon EFS caching tier that holds hot files for ~1ms access while passing large sequential reads and byte-range requests straight through to the underlying bucket. Supersedes the earlier **Mountpoint for S3** alpha (a FUSE-based read-mostly client) by moving the translation server-side and adding the EFS cache, write-back batching, and 25,000 concurrent connections per mount target.
For twenty years, workloads needing file semantics over S3 had to copy data into EBS or FSx and build brittle sync pipelines. S3 Files collapses that gap. The immediate beneficiaries are agentic AI systems — Claude Code, LangGraph orchestrators, multi-step model pipelines — which persist scratchpad memory and shared state using ordinary Python file I/O and expect edit-in-place semantics that a raw object store can't provide. AI training pipelines also benefit: legacy frameworks that assume POSIX (PyTorch DataLoader, NVIDIA DALI) can stream from S3 buckets directly, eliminating the dual S3+EFS topology and yielding **~90% cost reduction** versus running both services in parallel.
Agent scratchpad and shared memory, POSIX-bound analytics and scientific computing over lake data, lift-and-shift of legacy file-based apps onto S3, mixed-mode workloads that interleave NFS edits and REST reads on the same bucket.
Recent developments
- Independent stress testing surfaced real edge-case gaps. A stress test by The Register found that 6 of 10 deliberately awkward S3 objects silently vanished from the NFS mount with no client error, deleted files remained readable for 6–18 seconds after deletion, and objects created through the S3 API default to root:root 0644 ownership — causing permission conflicts when mounting through access points, where the lost+found directory is also invisible. Worth testing against your own key namespace before a production cutover. Per AWS put a file system on S3; I stress-tested it (The Register).
- Billing granularity matters more than headline rates. Every data access is metered at a 32 KB minimum, metadata operations are metered as a 4 KB read, and a commit (fsync or close-after-write) is metered as a 4 KB write — so small-file-heavy workloads can end up costing more than EFS despite the cheaper storage line. The offset: large sequential reads of 1 MB+ stream directly from the bucket with no S3 Files access charge at all, only standard S3 GET costs. Per S3 Is Not a Filesystem (Last Week in AWS) and AWS S3 Files vs EFS (AWS Fundamentals).
- Conflict losers land in lost+found. When the same file is modified simultaneously through NFS and the S3 API, S3 remains the source of truth — but the losing file-system version is moved to lost+found rather than silently discarded, leaving a recoverable trail for the "S3 wins" policy described above. Per Architecture Layers That S3 Files Eliminates — and Creates (dev.to).
Connections14
Outbound10
enables1constrained_by2Inbound4
consumes_via1alternative_to1competes_with1enables1Resources4
Official April 7 2026 launch notice for NFS v4.1 mount support on general-purpose S3 buckets.
Deep launch write-up covering the EFS caching layer, close-to-open vs atomic-PUT reconciliation, 60s write aggregation, and S3-wins conflict policy.
Werner Vogels' perspective on how S3 Files changes the 20-year separation between file and object storage — architectural framing worth reading before retrofitting.
Hands-on getting-started guide with mount examples, concurrency edge cases, and rename-cost discussion.