F3 is a future-oriented open source data file format, presented as a SIGMOD 2026 paper. It is released under the MIT License, with source code at github.com/future-file-format/F3.
Core Design Principles
- Efficiency — Vectorized decoding, with throughput better than Parquet.
- Interoperability — Self-describing files with native cross-platform compatibility.
- Extensibility — Generic APIs that support custom encodings.
- Old readers can still read files written with new encodings.
- Storage overhead is minimal (only KB-level).
- It fundamentally solves the format-evolution and version-compatibility problem that plagues existing columnar formats.
Key Innovation: Embedded Decoders (WebAssembly)
F3 files carry their own decoding logic as an embedded WebAssembly module layered alongside metadata and data. This means:
Architecture: IOUnit Decoupled Design
F3 decouples storage into IOUnits, removing Parquet's fixed row-group size limitation.
Performance Highlights (per the post)
| Metric | Result | |---|---| | Read throughput | Better than Parquet | | Random-access latency | Significantly lower | | Metadata parsing | 10x+ faster |
Summary
F3 combines vectorized decoding, self-description, embedded WASM decoders, and IOUnit-based storage to target next-generation storage workloads while remaining open source and backward compatible.