ZeroFS mounts S3 as a log-structured filesystem to bypass object latency

A technical review of ZeroFS, a new log-structured FUSE filesystem claiming major write performance gains over s3fs and goofys by treating S3 as sequential storage.

If you need to run legacy file-based applications directly on top of AWS S3 and your workload is write-heavy, ZeroFS is worth a look. It is designed for developers who are bottlenecked by the metadata overhead and random-write latency of traditional S3 FUSE mounts. However, if your workload requires multi-client concurrent writes or strict POSIX compliance for distributed systems, you should skip it. The bottom line is that ZeroFS trades standard object mapping for sequential log structures to achieve its speed.

Methodology

This v0 review draws on the creator's published claims at https://www.zerofs.net/ and the project's public repository as of July 2026. Independent benchmarks are currently pending. Our analysis covers the architectural design of ZeroFS, specifically how it structures data into sequential segments to minimize S3 API calls, and its positioning against established tools like s3fs and goofys. This review does not cover independent performance verification, long-term stability under heavy fragmentation, or edge-case handling of network partitions. We will update this review when our own test rig benchmarks diverge from the observed behavior reported by the creator, Eikon, who launched the project on Hacker News. We also examined the repository's documentation regarding metadata serialization and the garbage collection loop to assess the theoretical viability of the log-structured approach.

What it does

Log-structured write path

Unlike traditional FUSE filesystems that map each file directly to an S3 object, ZeroFS writes data sequentially to an append-only log. This design consolidates small, random writes into larger, sequential segments before flushing them to S3. This reduces the total number of PUT requests, which are both slow and costly on AWS.

Metadata caching and synchronization

ZeroFS maintains an in-memory metadata index to resolve file lookups instantly. Instead of querying S3 for every directory listing or file attribute, the system reads from its local index. This index is periodically serialized and saved to S3 as part of the log structure, allowing other read-only mounts to sync state.

Garbage collection and compaction

Because log-structured filesystems accumulate obsolete data during overwrites and deletions, ZeroFS includes a background compaction process. This cleaner identifies segments with high ratios of dead data, rewrites the active blocks into new segments, and deletes the old S3 objects to reclaim storage space.

Read path optimization

For reads, ZeroFS uses the in-memory index to locate the specific segments containing the requested file blocks. It then issues targeted range requests to S3. This prevents the need to download entire objects when only a fraction of the data is needed.

What's interesting and what's not

The architectural choice to use a log-structured layout on top of an object store is a compelling way to solve the write-latency problem. Traditional tools like s3fs attempt to maintain a one-to-one mapping between files and objects. This mapping forces s3fs to perform expensive multipart uploads or complete object rewrites for minor file updates. ZeroFS bypasses this by treating S3 as a block-like sequential tape.

The trade-off, however, is severe. Because files are fragmented across log segments, you lose the ability to read or manage your files directly through the AWS S3 console or other standard S3 clients. Your data is locked into the ZeroFS format. If the metadata index corrupts or the compaction process fails, data recovery becomes a complex forensic exercise.

Furthermore, the creator's claim of being "10x faster than s3fs" must be viewed with caution. This performance gain is highly dependent on write-heavy workloads with small files. For large, single-file sequential reads, the overhead of assembling fragmented blocks from multiple S3 objects could easily negate these benefits. The project also lacks a robust story for multi-client write concurrency, which is a common requirement for distributed cloud applications.

Pricing

ZeroFS is currently released as an open-source tool under the MIT License. There are no commercial tiers or licensing fees as of July 2026.

Verdict

ZeroFS is a highly specialized tool. It is an excellent choice if you have a single-writer, write-heavy legacy application that must use S3 for cheap, durable storage without rewriting the application code to use the AWS SDK. Do not use it if you need to access the underlying S3 objects directly from other AWS services like Athena or Lambda, as the log-structured format breaks standard object access. For general-purpose, multi-client read-write workloads, sticking to goofys or a managed service like AWS File Gateway remains the safer choice.

What we'd test next

In our next phase of testing, we plan to benchmark ZeroFS against s3fs and goofys using a standardized fio workload. We will measure write latency, CPU utilization, and AWS API call volume under three scenarios: a high-frequency small-file write loop, a large sequential file copy, and a sustained random-write test designed to trigger the background garbage collector. This will verify if the claimed performance holds up during active compaction.

The investor read

ZeroFS highlights a persistent friction point in cloud infrastructure: the mismatch between legacy POSIX file APIs and modern object storage. While AWS offers native solutions like EFS, their cost structures often drive developers back to S3-backed FUSE workarounds. ZeroFS's log-structured approach is technically elegant but remains a niche, developer-centric utility rather than a venture-scale platform. As an investable category, FUSE-based S3 adapters are highly vulnerable to cloud providers optimizing their own file-to-object gateways or lowering EFS pricing. However, watching how ZeroFS manages metadata consistency could signal design patterns for next-generation distributed databases that use S3 as a primary log store.

Pull quote: “Your data is locked into the ZeroFS format.”