export-from-scraper.py

#!/usr/bin/env python3
"""Export all trainview SQLite databases to a Hugging Face parquet dataset partitioned by year."""

import glob
import os

import duckdb

DB_DIR = os.path.join(os.path.dirname(__file__), "..", "website", "databases")
OUTPUT_DIR = os.path.join(os.path.dirname(__file__), "..", "huggingface")


def main():
    db_files = sorted(glob.glob(os.path.join(DB_DIR, "trainview-*.db")))
    if not db_files:
        raise SystemExit("No trainview-*.db files found")

    print(f"Found {len(db_files)} trainview databases")

    con = duckdb.connect()
    con.execute("INSTALL sqlite; LOAD sqlite;")
    con.execute("SET sqlite_all_varchar = true;")

    # ATTACH each SQLite file so DuckDB reads them directly (no Python materialization)
    union_parts = []
    for i, db_file in enumerate(db_files):
        alias = f"db{i}"
        con.execute(f"ATTACH '{db_file}' AS {alias} (TYPE sqlite);")
        union_parts.append(
            f"SELECT day AS service_date, time, train AS train_number, lateness "
            f"FROM {alias}.trainview"
        )

    union_query = " UNION ALL ".join(union_parts)

    print("Writing hive-partitioned parquet...")
    os.makedirs(OUTPUT_DIR, exist_ok=True)

    con.execute(f"""
        COPY (
            SELECT
                service_date::DATE AS service_date,
                time::VARCHAR AS time,
                train_number::VARCHAR AS train_number,
                lateness::INT32 AS lateness,
                YEAR(service_date::DATE) AS year
            FROM ({union_query})
            ORDER BY service_date, time, train_number
        )
        TO '{OUTPUT_DIR}'
        (FORMAT PARQUET,
         PARTITION_BY (year),
         OVERWRITE_OR_IGNORE,
         COMPRESSION 'zstd',
         ROW_GROUP_SIZE 100000);
    """)

    # Summary
    result = con.execute(f"""
        SELECT YEAR(service_date::DATE) AS year, COUNT(*) AS rows
        FROM ({union_query})
        GROUP BY year
        ORDER BY year
    """).fetchall()

    print("Rows per year:")
    for year, count in result:
        print(f"  {year}: {count:,}")

    con.close()
    print(f"\nParquet dataset written to {os.path.abspath(OUTPUT_DIR)}")


if __name__ == "__main__":
    main()