cozip 2026.5.20

Python bindings to libcozip, the reference library for the Cloud-Optimized ZIP (cozip) format.

cozip

Python bindings for libcozip. Read a Cloud-Optimized ZIP archive's manifest over HTTP range requests, or write one from a pyarrow Table.

The native libcozip binary ships inside the wheel, no C toolchain required.

Install

pip install cozip

Usage

Write

import cozip
import pyarrow as pa

table = pa.table({
    "name": ["a.txt", "b.bin"],
    "path": ["/path/to/a.txt", "/path/to/b.bin"],
})

cozip.write("out.zip", table)

name is how each file appears inside the archive. path is where it lives on disk, used at write time and dropped from the manifest. Any extra columns ride along into __metadata__ and become queryable on read.

table = pa.table({
    "name":      ["a.tif", "b.tif"],
    "path":      ["/path/to/a.tif", "/path/to/b.tif"],
    "cloud_pct": [12.3, 45.1],
})

cozip.write("out.zip", table)

Read

import cozip

manifest = cozip.read("https://example.com/dataset.zip")
train = manifest[manifest["split"] == "train"]

manifest is a pandas DataFrame with name, offset, size, cozip:gdal_vsi, plus whatever extras the writer added. Local file or remote URL, same call. Only the byte-0 index and the embedded __metadata__ Parquet are fetched, never the user payloads. Pass columns=[...] to bring only specific extras, gdal_vsi=False to drop the VSI path column.

Versioning

The Python package version tracks the C library, copied verbatim from the repo-root VERSION file at build time. The bundled .so / .dylib / .dll inside the wheel matches the wheel's version exactly, so there's no mismatch to worry about.

import cozip
cozip.__version__                  # "2026.5.16"
cozip.lib.cozip_version_string()   # full C library version

Spec

See SPEC.md for the on-disk format.

License

MIT. See LICENSE.