lbzip2 0.4.1

hmm # lbzip2-rs v0.4.0 "Skadi" ❄️

🧙‍♂ Med Allfaderns visdom och blick över varje bit.

Pure Rust parallel bzip2 decompressor. No C dependencies. Zero per-block allocations, physical-core aware, no polling. Usable as a library (in-process, zero-copy) or as a CLI tool:

# Decompress any bzip2 file (including pbzip2 concatenated streams)
cargo run --release --bin lbunzip2 -- planet-241021.osm.bz2 > planet.osm

What makes this crate unique: 100% Rust, zero-copy in-process library with a lightweight parallel split method that finds block boundaries with tiny forward-scans (~500 bytes per split). The decoder uses a streaming worker-pool (reorder buffer) to avoid global barriers and is tuned for modern CPU caches (packed Huffman tables fit L1, thread-local buffers leverage L2), yielding high throughput on many-core machines.

Part of the znippy group of software, designed for fast zero-copy integration with osm-katana — the parallel OSM-to-GeoParquet pipeline.

Performance vs C lbzip2

Laptop — T14s (Ryzen AI 360, 8 cores + SMT, 64 GB DDR5)

Planet 1 GB slice (1023 MB compressed → 9861 MB decompressed), interleaved runs, 20s cooldown:

Rust uses 12% less total CPU — the single-thread decode is 1.5× faster than C libbz2. On sustained workloads, thermal throttling on thin laptops can mask the advantage.

Single-thread decode (library, no I/O)

Odin — Threadripper PRO 3975WX, 32 cores, 512 GB DDR4

End-to-end: Planet bz2 → PBF (osm-katana)

Full pipeline: bz2 decompress → VTD XML parse → PBF encode, 15 workers.

Usage

use lbzip2::chunk::ChunkDecoder;

let data: &[u8] = /* compressed chunk including BZhN header */;
let decoder = ChunkDecoder::from_header(&data[..4])?;

// Returns segments separately — no giant memcpy
let (segments, consumed) = decoder.decode_chunk_segments(data, true)?;
for seg in &segments {
    // each seg is a Vec<u8> of decompressed data, in order
}

Single-stream sequential API:

let output = lbzip2::stream::decompress(&compressed)?;

Förklaring för Farbror Dan 🏍️🇸🇪

Hur fungerar worker-pool-arkitekturen?

Tänk dig 6 rum (slottar) som återanvänds i en ring. Varje rum rymmer 200 MB komprimerad data uppdelad i N väggar (en per CPU-kärna). N målare (workers) målar väggar utan att stanna — när en vägg i rum 1 är klar går målaren direkt till rum 2.

Rum 1:  [████████████░░] Lisa nästan klar
Rum 2:  [██████░░░░░░░░] Kalle + andra redan igång!
Rum 3:  [██░░░░░░░░░░░░] snabbaste redan här

Ingen barriär, ingen väntan. Materialaren fyller rum, målarna målar, skrivaren skriver — som ett löpande band. 🐴

License

MIT OR Apache-2.0, plus the original bzip2 license (BSD-style, Julian Seward) for the block-decode routines inspired by the C reference implementation. See LICENSE-BZIP2 for the full text.