Convert Parquet to CSV from the Command Line

The one-liner that usually wins: DuckDB

The DuckDB CLI↗ is a single dependency-free binary, which makes it the cleanest command-line route:

duckdb -c "COPY (SELECT * FROM 'events.parquet') TO 'events.csv' (HEADER, DELIMITER ',');"

That streams the file — it never loads the whole dataset into memory — and it auto-detects Snappy, Gzip, ZSTD and LZ4 compression from the file metadata. Because the source is just a SELECT, you can trim during the conversion instead of exporting everything and cleaning up afterwards:

duckdb -c "COPY (
  SELECT user_id, event_type, created_at
  FROM 'events.parquet'
  WHERE created_at >= '2026-08-01'
) TO 'august.csv' (HEADER);"

Two more variants worth knowing:

# Many files -> one CSV (single header row)
duckdb -c "COPY (SELECT * FROM 'exports/*.parquet') TO 'all.csv' (HEADER);"

# Straight to stdout, for piping into grep/awk/another tool
duckdb -csv -c "SELECT * FROM 'events.parquet'" | head -20

For a semicolon-delimited file (common for European Excel locales), set DELIMITER ';' in the COPY options.

Python: pandas or pyarrow

If a Python environment is already part of the pipeline, pandas is the familiar route:

python -c "import pandas as pd; pd.read_parquet('events.parquet').to_csv('events.csv', index=False)"

Note index=False — without it pandas prepends an unnamed index column that confuses every downstream consumer. For large files, pyarrow's CSV writer skips the pandas conversion step and is noticeably faster:

import pyarrow.parquet as pq
import pyarrow.csv as pcsv

pcsv.write_csv(pq.read_table("events.parquet"), "events.csv")

Both approaches load the table into memory, so for files in the multi-GB range prefer the DuckDB streaming route above.

No terminal at all: convert in the browser

When the file is on a machine where you cannot (or do not want to) install anything — a locked-down work laptop, someone else's desk — the Parquet to CSV converter does the same job in a browser tab. It runs DuckDB compiled to WebAssembly locally, so the file is never uploaded; that distinction matters when the data is confidential. If you want to filter or sample before exporting, the SQL Workbench accepts the same file and lets you download any query result as CSV.

What to check after converting

CSV discards Parquet's type system, so a quick sanity pass saves pain downstream:

• Timestamps become plain strings. DuckDB writes ISO 8601, which most tools re-parse cleanly; pandas output follows its own formatting.
• Decimals and big integers may lose their declared precision — Excel in particular will happily round an 18-digit ID.
• Nested columns (lists, structs) are serialized as text in a single cell. If you need the structure, export JSON or JSONL instead.
• Row count: compare wc -l output.csv (minus the header) against the source. The row count is in the Parquet footer — drop the file into the Parquet Viewer to read it without scanning any data.