UC2 for Python

Read and create UltraCompressor II archives using the native UC2 library. The package also includes the UC2 command-line program for listing, extraction, archive creation, and stream ingestion. Wheels contain the native code; no separate UC2 install or compiler is needed when a wheel is available.

This is an alpha release. It supports the implemented legacy Huffman methods, master dictionaries, directories, stored versions, and UC2 v3 rANS entries. Turbo/method 80 and multi-volume archives remain unsupported. Use UC2File for reading and UC2Writer for creation.

Install and read

python -m pip install --pre uc2
uc2 -l archive.uc2
python -m uc2 -t archive.uc2

Wheels support CPython 3.10-3.14 on Linux x86-64 (glibc), Windows x86-64, and Intel/Apple Silicon macOS.

from uc2 import UC2File

with UC2File("archive.uc2") as archive:
    for info in archive.infolist():
        print(info.filename, info.file_size, info.version)
    contents = archive.read("notes.txt")
    archive.verify()
    archive.extractall("restored")

The constructor accepts a path or a seekable binary file, including BytesIO. Caller-owned files stay open; their seek position changes. Keep the source unchanged while the archive is open. Each thread should use its own UC2File.

infolist() and namelist() return the newest version of each path by default. Use infolist(all_versions=True) to include older entries. A literal semicolon in a filename remains part of the name:

with UC2File("archive.uc2") as archive:
    newest = archive.read("notes.txt")
    older = archive.read("notes.txt", version=1)
    literal_name = archive.read("notes.txt;1")

UC2Info records the name, uncompressed and compressed sizes, DOS timestamp (a naive datetime, or None if invalid), attributes, method, version, and is_dir(). Metadata objects can be passed directly to read, extract, and copy_to. They belong to the archive that produced them.

Limits and extraction

Defaults are 100,000 entries, 64 MiB of decoder allocations, 64 MiB per read, 1 GiB per streamed file, and 1 GiB total per extractall or verify call. Constructor keywords are max_entries, max_decoder_memory, max_read_size, max_extract_size, and max_total_size. They must be positive integers. The decoder limit excludes Python metadata and caller-owned buffers. Archive offsets and this reader are limited to 32 bits. These limits are not a CPU-time sandbox for hostile input.

copy_to(member, binary_file) decodes without collecting the file in memory. Its output may be partial if decoding or writing fails. extract and extractall instead stage each file and install it after its UC2 checksum passes. Existing files are preserved unless overwrite=True is supplied. Earlier successfully extracted files remain if a later member fails.

Extraction rejects absolute paths, traversal, reserved Windows names, control characters, ambiguous case-folded names, and links below the destination. It uses file-descriptor-relative operations without following links on POSIX. Windows checks links and reparse points; use a destination that other processes cannot modify during extraction. The explicitly chosen destination itself may be a symlink. POSIX files use mode 0600 and directories use 0755, both masked by the process umask. Archived permissions and timestamps are exposed as metadata but not applied.

Errors include BadArchiveError, ResourceLimitError, UnsafePathError, and ordinary Python I/O exceptions. Verification uses the archive's legacy checksums, which are not cryptographic authentication. A missing entry or version raises KeyError.

Create archives

from io import BytesIO
from uc2 import UC2Writer

with UC2Writer("backup.uc2", level=4) as archive:
    archive.write("project", arcname="source")  # recursive directory tree
    archive.mkdir("empty-directory")
    archive.writestr("notes.txt", "Saved from Python\n".encode("utf-8"))
    archive.write_stream("data.bin", BytesIO(b"streamed bytes"))

The destination is a filesystem path. A successful close() replaces it using an exclusively created sibling temporary file and os.replace. An exception, abort(), or a failed add/close leaves an existing destination intact. Any failed add aborts the writer; start a new one to retry. Destruction discards uncommitted work. The containing directory must exist and remain under the caller's control. Atomic replacement is not a guarantee of survival after a power failure. Replacement creates a new file with temporary-file permissions (normally 0600 on POSIX), without preserving destination permissions.

write(path, arcname=None) adds a regular file or recursively adds a directory. The default archive name is the input's basename. Symlinks, reparse points, and special files are rejected. Keep the source tree unchanged while reading. mkdir(name) creates missing parents; repeated directories retain their first metadata. writestr(name, data) accepts bytes, bytearray, or memoryview; encode text explicitly. write_stream(name, source, size=None) consumes read(n) through EOF without seeking and leaves the input open. A supplied size must match exactly; detecting excess input can consume one extra byte.

Inputs spool to anonymous temporary files as they are added. Compression runs at close, using CDC-selected custom master dictionaries. Payload buffers have fixed sizes; metadata grows with entries and chunks. Temporary disk space covers the uncompressed inputs and completed archive, plus the destination's sibling temporary file during the final copy. Duplicate file paths store versions: the last added version is newest. Directories cannot conflict with files. Independent writer objects may run separately; one object rejects concurrent or reentrant use.

Writer limits and metadata:

• level=4 by default; 2-5 use legacy Huffman and 6-9 use v3 rANS.
• max_entries=100_000, configurable from 1 through 1,000,000, includes automatic parent directories. max_total_size=0xFFFFFFFF bounds all input bytes combined and may be reduced. Each file and the completed archive also fit 32 bits.
• The central directory is limited to 64 MiB and the CDC index to 1,048,576 chunks. Directories may nest 64 levels deep. These are format/API limits, not CPU limits.
• Archive paths use /. Components must be representable in CP850, fit 255 CP850 bytes and 299 UTF-8 bytes, and contain no traversal, separators, colon, or control characters. Unsupported Unicode raises an error; names are never shortened. Some legal archive names still require renaming to extract on Windows; the reader's extraction checks apply unchanged.
• mkdir, writestr, and write_stream accept date_time (a naive datetime in 1980-2107, or None) and DOS attributes (R/H/S/A bits). Timestamps have two-second resolution. write takes filesystem modification times; values outside that range become unspecified. It stores ordinary file/directory attributes, not Unix permissions.

After successful close, writer.stats reports file, directory and master counts and archive size. The C interface is documented in uc2_writer.h and the library guide.

Read-only Linux mount

python -m pip install --pre 'uc2[mount]'
mkdir archived
uc2-mount archive.uc2 archived --cache-mib 512

The mount runs in the foreground. Use another terminal to browse archived/files and archived/versions/1, then unmount with fusermount -u archived. Linux libfuse2 and /dev/fuse are required. Files decode on first access into a checksum-verified disk cache; an entry larger than its budget fails with EFBIG. The default cache holds up to 512 MiB and 128 files. Decoder allocations are limited to 64 MiB and released after every decode. Python metadata is outside that budget and has separate entry and node limits. See the mount documentation for options, stored versions, and error behavior.

Build from source

Requires Python 3.10+, CMake, and a C99 compiler. The build backend obtains CMake when necessary.

python -m pip install .
python -m build

Source, releases, and the project roadmap are at github.com/evvaletov/uc2.

The distribution contains GPL-3.0-or-later and LGPL-3.0-only code. Independently written Python integration code also has a BSD-2-Clause option. This does not change the licenses of the bundled core or third-party code. License texts, credits, and the per-file audit are included in the distribution.