Researchers behind F3 have released an open-source prototype of a data file format intended to improve efficiency and extensibility beyond established formats such as Parquet. The project accompanied work presented at SIGMOD 2026 and uses embedded WebAssembly decoders to support user-defined encodings without requiring every reader to understand them in advance.
The repository describes F3 around three goals: efficient data organization, interoperability and the ability to evolve. Its designers argue that older columnar formats contain layout limitations that are difficult to correct while maintaining compatibility. F3’s proposed response is to define a new organization and let a file carry decoding logic for additional encodings through WebAssembly.
That architecture aims to separate the format’s durable container from the pace of encoding experimentation. A reader with a WebAssembly runtime could execute an embedded decoder even when the encoding was introduced after the reader was built. In principle, this can make a format more future-proof and reduce the need to coordinate native decoder releases across a broad tool ecosystem. The supplied evidence does not provide security analysis, however, and executing logic associated with a data file introduces trust, sandboxing and resource-governance questions that production systems would need to address.
The codebase includes a FlatBuffers definition of the format, the main proof-of-concept implementation, core and encoding components, and modules for user-defined encoding and WebAssembly decoding. Separate benchmark code contains the microbenchmarks and end-to-end experiments reported in the paper. Scripts and documentation provide reproduction steps, allowing other researchers to inspect the implementation and rerun the published experiments.
The maintainers explicitly label F3 a research prototype and advise against production deployment. Testing was limited to an Intel machine running Debian 12. That warning is material: it means compatibility across architectures and operating systems, resilience against malformed inputs, upgrade behavior and operational performance have not been established for general workloads.
F3 is available under the MIT licence. The open repository and reproduction guidance make the immediate audience database researchers and systems engineers evaluating the paper’s claims, rather than application teams choosing a stable storage standard.
A new file format succeeds only partly on its technical design. It also needs reliable implementations, language bindings, query-engine integration, long-term governance and adoption sufficient to justify migration. F3’s embedded decoder concept offers one answer to the extensibility problem, but its practical tradeoffs will become clearer through independent benchmarks and security review. For now, the release is best understood as an inspectable experiment in how analytical data files might evolve, not as a drop-in replacement for Parquet. That staged positioning is appropriate for systems research.
