geoparquet-testing

Official test corpus for the GeoParquet format. Modeled after apache/parquet-testing.

Targets: GeoParquet 2.0.0 (version: "2.0.0" in geo metadata). Corpus version: see CHANGELOG.md and releases.

Each tier directory has its own README indexing every file and what it tests.

The corpus targets GeoParquet 2.0.0 and every file in data/ and samples/ validates against the GeoParquet 2.0.0 JSON Schema (CI enforces this). Every geometry column uses the native Parquet GEOMETRY/GEOGRAPHY logical type, and every column whose CRS is not OGC:CRS84 carries that CRS on the native logical type (stamped as inline PROJJSON via sedonadb — see scripts/gen_native_crs.py) so it agrees with the geo metadata, as the spec requires.

Two niche CRS-representation variants remain intentionally deferred (tooling limits):

• srid:0 in the Parquet metadata + null geo crs (an unknown CRS), and
• full PROJJSON in the geo crs + the compact authority:code string in the Parquet native metadata (our tooling only writes inline PROJJSON onto the native type).

git submodule add https://github.com/geoparquet/geoparquet-testing tests/data/geoparquet-testing

Pin to a tagged release by checking out the tag inside the submodule.

Download from the releases page. Suitable for projects that don't want a submodule dependency.

A reader implementation can iterate bad_data/manifest.json, attempt to open each file, and assert that the failure mode matches the manifest's expected_failure code. See bad_data/README.md for the controlled vocabulary.

cd scripts
uv sync --all-extras
uv run python generate_all.py

All conformance (data/) and negative (bad_data/) files are deterministic — re-running produces byte-identical output, and CI enforces this on every PR. Realistic samples/ that fetch live upstream data are committed snapshots (not byte-diffed in CI).

To regenerate a single category:

uv run python gen_crs.py
uv run python gen_samples.py --only airports_global

1. Edit or add a generator under scripts/.
2. Run uv run python generate_all.py.
3. Make sure the relevant data|samples|bad_data/README.md table includes any new file.
4. For new bad-data files, add an entry to bad_data/manifest.json (the generator does this automatically when you register(...)).
5. Open a PR. CI verifies reproducibility (data/ + bad_data/), schema validity, README index hygiene, and the 5 MB sample budget.

Apache 2.0 for everything generated by this repo. Realistic samples under samples/ retain their upstream licenses — see each entry in samples/README.md and the header comment of the matching scripts/samples/<name>.py for attribution requirements.