Example files for GEOMETRY and GEOGRAPHY logical type

[paleolimbot]

As discussed on the mailing list, it's best to get example files early!

Code to generate in details (requires https://github.com/apache/arrow/compare/main...paleolimbot:arrow:parquet-geo-write-files-from-geoarrow , which is a slightly more functional but less appropriate initial version of https://github.com/apache/arrow/pull/45459 ). I've also added the full suite of geoarrow-data files (even the big ones) to that forthcoming release: https://github.com/geoarrow/geoarrow-data .

import urllib.request
import json

import pyarrow as pa
from pyarrow import parquet
import geoarrow.pyarrow as ga

manifest_url = (
    "https://raw.githubusercontent.com/geoarrow/geoarrow-data/v0.2.0-rc4/manifest.json"
)
files = {}
with urllib.request.urlopen(manifest_url) as f:
    manifest = json.load(f)
    for group in manifest["groups"]:
        for file in group["files"]:
            if file["format"] == "arrows/wkb":
                files[group["name"] + "_" + file["name"]] = file["url"]

out_dir = "/Users/dewey/gh/parquet-testing/data/geospatial"
ones_that_didnt_work = []
for name, url in files.items():
    # Skip big files + one CRS example that includes a non-PROJJSON value
    # on purpose (allowed in GeoArrow), which is rightly rejected
    # by Parquet
    if (
        "microsoft-buildings" in name
        or ("ns-water" in name and name != "ns-water_water-point")
        or "wkt2" in name
    ):
        print(f"Skipping {name}")
        continue

    # Maintain chunking from IPC into Parquet
    out = f"{out_dir}/{name}.parquet"
    with (
        urllib.request.urlopen(url) as f,
        pa.ipc.open_stream(f) as reader,
        parquet.ParquetWriter(
            out,
            reader.schema,
            store_schema=False,
            compression="none",
            write_geospatial_logical_types=True,
        ) as writer,
    ):
        original_schema = reader.schema
        print(f"Reading {url}")
        for batch in reader:
            writer.write_batch(batch)
        print(f"Wrote {out}")
    
    # Read in original table for comparison
    with (
        urllib.request.urlopen(url) as f,
        pa.ipc.open_stream(f) as reader
    ):
        original_table = reader.read_all()

    print(f"Checking {out}")
    with parquet.ParquetFile(out, arrow_extensions_enabled=True) as f:
        if f.schema_arrow != original_table.schema:
            print(f"Schema mismatch:\n{f.schema_arrow}\nvs\n{original_schema}")
            continue

        reread = f.read()
        if reread != original_table:
            print("Table mismatch")

[paleolimbot]

@Kontinuation @zhangfengcdt Can you give these a try from Java when you're ready? I'm fairly confident that they are correct, including the "crs" examples that dump the actual payload of the PROJJSON to the file metadata.

[paleolimbot]

I pushed an update to three files here - the original fields that PROJJSON crses were written to were very likely to collide with eachother if you did things like read a Parquet file, filter it, then write it again 😬 . The new files add a hash of the value to the end of the key (e.g., projjson_crs_value_0ffad8372). Totally up for discussion whether that's a good idea or not 🙂 .

[paleolimbot]

I updated these to be a bit more intentional about the corner cases we collectively ran into in https://github.com/apache/parquet-java/pull/2971 and https://github.com/apache/arrow/pull/45459. I'm not sure the Python files to generate them belong in this repo but it does make it easier to see what they contain. I also included CRS examples because that was also something that required some thinking about in the C++ PR...happy to remove or tweak any of these if I didn't get the spirit of the format change right 🙂 .

[alamb]

Today at the Parquet sync @emkornfield said he might have some time to review this PR

[emkornfield]

This all seems reasonable, going to merge.

[alamb]

Thank you @emkornfield and @paleolimbot 🙏

[paleolimbot]

Thank you both!