Retrieve a file from a 5 GB archive in 7 ms. 811x faster than tar.gz.
ARCX is a compressed archive format built for modern workflows. It lets CI/CD systems, build tools, and cloud storage retrieve one file without unpacking the whole archive — locally or remotely via HTTP range requests.
Fast to create. Fast to query. Works with S3 and HTTPS. No full decompression required.
Demo: extracting one file from an 800 MB archive. TAR+ZSTD: 239 ms. ARCX: 13 ms.
Full details: Benchmark methodology | Format specification
Traditional archives force a choice: compress well (tar+zstd) or access fast (zip). ARCX does both.
- Cross-file compression matching tar+zstd ratios
- Indexed, block-level access to any file
- Single-digit millisecond retrieval (7 ms from a 5 GB archive)
- Remote extraction via HTTP range requests (S3, HTTPS)
- Up to 200x less data read than tar+zstd
- Python bindings:
pip install getarcx - GitHub Action:
getarcx/cache@v1
ARCX reads kilobytes instead of megabytes.
- ZIP gives fast per-file access, but weak cross-file compression.
- TAR+ZSTD gives strong compression, but slow selective access.
- ARCX is designed to deliver both.
# Install (Rust)
cargo install arcx
# Or Python
pip install getarcx
# Pack a directory
arcx pack ./my-project output.arcx
# Get a single file (instant)
arcx get output.arcx src/main.rs --time
# List contents
arcx list output.arcx
# Extract everything
arcx extract output.arcx ./output-dirExtract files from remote archives without downloading the whole thing. ARCX uses HTTP range requests to fetch only the footer, manifest, and needed blocks.
# Extract one file from a remote archive (downloads only ~3% of the archive)
arcx get https://example.com/artifacts.arcx path/to/file.txt
# Works with S3 URLs
arcx get s3://my-bucket/build.arcx src/main.rs
# List files remotely (downloads only the manifest — 0.09% of archive)
arcx list s3://my-bucket/build.arcx
# Remote archive info
arcx info https://example.com/artifacts.arcxRemote benchmark (15 MB archive over HTTPS):
| File | Downloaded | % of Archive | HTTP Requests | Time |
|---|---|---|---|---|
| 3.6 KB source file | 455.8 KB | 3.01% | 3 | 267 ms |
| 81 KB data file | 94.9 KB | 0.63% | 3 | 314 ms |
| 3.6 MB asset | 3.6 MB | 24.28% | 6 | 730 ms |
| List all 128 files | 13.8 KB | 0.09% | 2 | — |
Drop-in replacement for actions/cache using ARCX for faster selective restoration:
- uses: getarcx/cache@v1
with:
path: build/artifacts
key: build-${{ hashFiles('**/Cargo.lock') }}How much data does the format actually read from disk to extract ONE file?
ARCX reduces unnecessary data transfer by up to 200x.
| Dataset | Target File | ARCX | TAR+ZSTD | ZIP |
|---|---|---|---|---|
| Build artifacts (581 files, 181 MB) | 36.4 KB .o file | 713.8 KB | 140.4 MB | 36.5 KB |
| Python ML project (206 files, 64 MB) | 6.9 KB .py file | 326.1 KB | 63.1 MB | 2.1 KB |
| Log archive (1,008 files, 30 MB) | 35.8 KB log file | 365.8 KB | 5.0 MB | 6.1 KB |
| Node.js project (19,001 files, 43 MB) | 1.5 KB .d.ts file | 1.3 MB | 17.4 MB | 632 B |
| Source code repo (389 files, 1.9 MB) | 4.5 KB .go file | 373.8 KB | 656.1 KB | 1.9 KB |
ARCX reads the manifest plus one compressed block. TAR+ZSTD reads the entire archive. ZIP reads only the individual file entry (no cross-file compression to navigate).
The defining metric. How long to extract ONE file from a sealed archive?
Build Artifacts (581 files, 181 MB)
ARCX |## | 7.9 ms
ZIP |# | 1.7 ms
TAR+ZSTD |######################### | 147.7 ms
TAR+GZ |################################################ 409.3 ms
Log Archive (1,008 files, 30 MB)
ARCX |# | 7.3 ms
ZIP |# | 3.0 ms
TAR+ZSTD |############ | 99.5 ms
TAR+GZ |################################################ 238.4 ms
Python ML Project (206 files, 64 MB)
ARCX |# | 2.8 ms
ZIP |# | 0.9 ms
TAR+ZSTD |############### | 53.4 ms
TAR+GZ |################################################ 180.0 ms
Node.js Project (19,001 files, 43 MB)
ARCX |##### | 131.3 ms
ZIP |## | 62.5 ms
TAR+ZSTD |################################ | 840.9 ms
TAR+GZ |################################################ 1.26 s
Large Benchmark (760 files, 5 GB)
ARCX |# | 7.3 ms
TAR+GZ |################################################ 5.68 s 811x slower
ZIP also supports random access via its central directory -- that is why it appears fast in these charts. The comparison is most meaningful against TAR-based formats, which represent the majority of compressed archives in CI/CD and cloud storage. ARCX combines the compression advantage of TAR+ZSTD with direct file access.
ARCX makes archive access effectively direct and block-addressable.
All measurements are cold-cache medians (3 runs each) on Windows 11, Intel i7 (22 cores). Datasets are synthetic but representative of real-world workloads. See benchmarks/ for full methodology and raw data.
ARCX matches TAR+ZSTD compression ratios across all workloads. Both outperform ZIP due to cross-file compression.
| Dataset | Files | Input | ARCX | TAR+ZSTD | ZIP |
|---|---|---|---|---|---|
| Node.js project | 19,001 | 42.6 MB | 18.8 MB (44.0%) | 17.4 MB (40.8%) | 21.7 MB (50.9%) |
| Python ML project | 206 | 63.8 MB | 63.1 MB (98.9%) | 63.1 MB (98.9%) | 63.1 MB (98.9%) |
| Build artifacts | 581 | 180.9 MB | 140.5 MB (77.6%) | 140.4 MB (77.6%) | 140.5 MB (77.7%) |
| Log archive | 1,008 | 30.4 MB | 5.0 MB (16.5%) | 5.0 MB (16.6%) | 5.1 MB (16.8%) |
| Source code repo | 389 | 1.9 MB | 683.5 KB (34.3%) | 656.1 KB (33.0%) | 721.8 KB (36.3%) |
ARCX groups files into compressed blocks. Multiple small files share a single block. Large files span multiple blocks. Each block is independently decompressible.
+----------+-------+-------+-------+-----+----------+--------+
| Header | Block | Block | Block | ... | Manifest | Footer |
| (80 B) | 0 | 1 | 2 | | | (40 B) |
+----------+-------+-------+-------+-----+----------+--------+
The manifest is a binary index stored at the end of the archive. It maps every file path to the block(s) containing its data, along with offsets and sizes. The manifest itself is zstd-compressed and uses delta encoding and a string table to minimize overhead.
On a get operation, ARCX:
- Reads the 40-byte footer to locate the manifest
- Reads and decompresses the manifest
- Looks up the target file's block reference(s)
- Seeks to and decompresses only the relevant block(s)
- Extracts the file data from the decompressed block
Total I/O: footer + manifest + one block. Everything else is untouched.
For remote archives (S3/HTTPS), the same logic applies over HTTP range requests. ARCX fetches the footer (40 bytes), manifest (~14 KB), and only the needed block(s) — typically downloading less than 3% of the archive.
Because multiple files are packed into the same block before compression, zstd sees cross-file redundancy. This is why ARCX matches TAR+ZSTD compression ratios -- both compress files together -- while ZIP compresses each file independently.
- CI/CD artifacts -- Store build outputs once, retrieve individual files on demand without downloading and decompressing the full archive.
- Cloud storage -- Reduce egress costs by reading only the bytes you need. Combine with HTTP range requests for remote selective extraction.
- Package managers -- Inspect package metadata or extract a single file without pulling the entire package.
- Game assets -- Load individual textures, models, or audio files from a compressed asset bundle at runtime.
- Log archives -- Query a specific day's logs from a compressed monthly archive without decompressing the other 29 days.
Block-based compression and indexed access exist in specialized systems (e.g., BGZF in bioinformatics, WARC for web archives). ARCX brings these ideas into a general-purpose archive format designed for modern developer workflows.
- Manifest index overhead is still being optimized for archives with 100K+ files.
- Current benchmarks focus on selective access; full extraction performance parity with tar+zstd is expected in the Rust implementation but not yet benchmarked at scale.
See FORMAT.md for the complete binary format specification.
Header (80 bytes)
Magic, version, offsets to manifest and data, flags, timestamp
Blocks (variable)
4-byte LE size prefix + zstd-compressed payload
Each block contains one or more file chunks
Manifest (variable)
Binary index: string table, file entries, chunk entries, block entries
Delta-encoded offsets, varint sizes, zstd-compressed
Footer (40 bytes)
Magic, manifest offset/size, total file count, checksum
git clone https://github.com/getarcx/arcx.git
cd arcx
cargo build --releaseThe binary will be at target/release/arcx (or arcx.exe on Windows).
Licensed under the Apache License, Version 2.0. See LICENSE for details.
Built by Highpass Studio
