fc is a single-file C library that compresses IEEE-754 doubles by running ~50 specialized codecs per block and keeping the smallest result.
Key Takeaways
Achieves 3.07x compression ratio on 17 synthetic double datasets vs. zstd-3 at 2.07x; wins ratio on 10 of 17 datasets outright.
Encode throughput is ~120 MB/s (multi-threaded, AVX2); decode hits ~1.28 GB/s, making it suited for write-once/read-many time-series stores.
Loses to zstd on quantized or dictionary-friendly data (e.g., decimal-cents: zstd-9 3,465x vs. fc 268x) and to fpzip on noisy scientific arrays by small margins.
Hard requirements: x86-64 with AVX2 + SSE4.2 + BMI + LZCNT; input must be 8-byte-aligned multiples of 8 bytes; no ARM/NEON path.
On-disk format is versioned by magic number but not stable across major versions; unknown mode IDs silently decode to zeros – treat untrusted streams carefully.
Hacker News Comment Review
The author clarifies fc is narrowly scoped: float-specific predictors and transforms for time-series, scientific, simulation, and analytics data, not a general zstd/lz4 replacement.
A commenter asked whether fc has been compared to Chimp128 or Arrow’s byte stream split encoding, two common alternatives in columnar float compression pipelines – no response yet.
Notable Comments
@Scaevolus: asks about Chimp128 and Arrow byte stream split as missing baselines in the benchmark.
