Batch Convert a Folder of Parquet Files to CSV

The situation

A pipeline dropped forty Parquet files into exports/ — one per day, partition or job run — and the downstream consumer wants CSV. Not one big CSV, but 2026-10-01.csv, 2026-10-02.csv and so on, one per input. This is a loop problem, not a SQL problem, and it takes one line in most shells.

DuckDB in a shell loop (macOS / Linux)

DuckDB is a single binary, streams each file, and handles every common Parquet codec. Loop over the files and run one COPY per file:

mkdir -p csv
for f in exports/*.parquet; do
  name=$(basename "$f" .parquet)
  duckdb -c "COPY (SELECT * FROM '$f') TO 'csv/$name.csv' (HEADER, DELIMITER ',')"
done

Each run starts a fresh in-memory DuckDB, so a 5 GB file in the middle of the batch does not hold memory for the rest of the loop. CSV options go inside the COPY parentheses: HEADER, DELIMITER ';', NULLSTR ''.

PowerShell (Windows)

New-Item -ItemType Directory -Force csv | Out-Null
Get-ChildItem exports\*.parquet | ForEach-Object {
  $out = "csv\$($_.BaseName).csv"
  duckdb -c "COPY (SELECT * FROM '$($_.FullName)') TO '$out' (HEADER)"
}

pyarrow, streaming per batch

If DuckDB is not available but pyarrow is, convert batch by batch so no file is ever fully loaded:

from pathlib import Path
import pyarrow.parquet as pq
import pyarrow.csv as pv

out_dir = Path("csv")
out_dir.mkdir(exist_ok=True)

for src in sorted(Path("exports").glob("*.parquet")):
    pf = pq.ParquetFile(src)
    with pv.CSVWriter(out_dir / f"{src.stem}.csv", pf.schema_arrow) as writer:
        for batch in pf.iter_batches(batch_size=100_000):
            writer.write_batch(batch)
    print(f"{src.name}: {pf.metadata.num_rows} rows")

Avoid the pd.read_parquet(f).to_csv(...) pattern for large batches — it loads each whole file into memory, and nulls in integer columns turn into floats (42.0) in the output.

In the browser, no install

The Parquet to CSV converter accepts multiple files in one drop. Select every file in the folder, and each is converted locally with DuckDB WebAssembly into a CSV named after its source (2026-10-01.parquet becomes 2026-10-01.csv). A Download all button saves the batch. Nothing is uploaded, which matters when the folder holds customer data.

Check the batch before you ship it

Two quick checks catch most problems:

• Row counts match. Compare each input with SELECT count(*) FROM 'exports/2026-10-01.parquet' against wc -l on the CSV (minus one for the header). A mismatch usually means a string column contains embedded newlines — still valid quoted CSV, but worth knowing before someone splits it by line.
• Schemas agree. Count schema entries per file in the SQL Workbench or the CLI: SELECT file_name, count(*) FROM parquet_schema('exports/*.parquet') GROUP BY ALL ORDER BY 2. A file with a different count will produce a CSV with different headers.

If what you actually need is a single file, read the folder with a glob and write once — see merging multiple Parquet files. For null strings, date formats and quoting, see Parquet to CSV formatting options.