split_by_domain.py

#!/usr/bin/env python3

from pathlib import Path
import pyarrow as pa
import pyarrow.parquet as pq

# Ordner
INPUT_DIR = Path("data")
OUTPUT_DIR = Path("layers/domain")
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

writers = {}

for parquet_file in sorted(INPUT_DIR.glob("*.parquet")):
    print(f"→ {parquet_file.name}")

    table = pq.read_table(parquet_file)

    domains = table.column("domain").to_pylist()

    for domain in ("code", "math", "science"):

        mask = pa.array([d == domain for d in domains])

        subset = table.filter(mask)

        if subset.num_rows == 0:
            continue

        outfile = OUTPUT_DIR / f"{domain}.parquet"

        if domain not in writers:
            writers[domain] = pq.ParquetWriter(
                outfile,
                subset.schema,
                compression="zstd"
            )

        writers[domain].write_table(subset)

for writer in writers.values():
    writer.close()

print("\n✅ Fertig.")