Answers are grounded in this paper and cite their sources.
Evaluates a novel filesystem substrate, improving synchronization and versioning in cloud environments, indicating potential broad applications.
Parallel software agents increasingly edit one body of files from different processes and machines. They need Git-like retained history and conflict visibility, but synchronization must be continuous rather than organized around explicit repository commits. Versioning, synchronization, encrypted remote storage, and portable execution are commonly separate layers whose boundaries give one update several identities: a filesystem block, a version-control delta, a synchronization object, and a cloud object. We present sfs, a filesystem substrate motivated by this agentic-time setting that instead makes a position-addressed fragment version map the common representation. Each unit carries a sparse version vector and, for every fragment position, the causal dot of the bytes currently selected there. The map is simultaneously a retained local lineage, a synchronization manifest, and the basis for sub-file conflict classification. Concurrent updates to disjoint fragments merge without a content-specific CRDT. Concurrent updates to the same fragment remain visible as separate strains. Around this representation, sfs combines deterministic fixed-size chunking, atomic root publication, retained versions and named cuts, client-side-encrypted synchronization, and an epoch-versioned multi-writer authorization model. The same synchronization protocol operates against either a blind cloud store or a live peer. In the secure deployment profile, content and metadata are client-side AES-GCM encrypted, and accepted record frontiers are authorized by an owner-signed writer set. The service receives neither plaintext nor the keys needed to derive it. A single v12 container format is implemented by a memory-safe Rust core, a FUSE adapter over host-file containers, an independent in-kernel C driver over raw devices, a C ABI, and experimental in-memory WebAssembly bindings. The same logical state therefore crosses an execution boundary (kernel, userspace, or browser) and a trust boundary (local media or encrypted service) without conversion into a second storage model. Deterministic multi-replica scenarios exercise disjoint-fragment convergence, same-fragment strain preservation and resolution, third-replica discovery, and bidirectional synchronization with a live TLS peer. They establish the mechanisms in small controlled topologies, not behavior at agent-fleet scale. We separately evaluate the mounted Linux paths over 14 I/O sizes using ten fault-checked runs per cell. Under the reported single-threaded protocol, the kernel-path geometric mean is 1.22× the ext4 baseline for plaintext writes and 0.98× for plaintext cold reads. Against ext4-on-LUKS, it is 0.98 to 0.99× for encrypted writes and 1.11 to 1.20× for encrypted cold reads. The FUSE path reaches 3.11× the fuse2fs baseline for plaintext writes and 1.70× for plaintext cold reads, but only 0.33 to 0.38× the gocryptfs baseline for encrypted cold reads. These results establish a competitive native path and identify the userspace encrypted-read path as the principal performance limit. They do not constitute a general workload or scalability evaluation. As a secondary result, an annex reports a NoSQL projection built without changing the core representation.
No takes yet. Share an insight, caveat, or question.
Sandra Theresa Keßler (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: