CSV vs Parquet: size, speed and when to switch

CSV

CSV is plain text with one record per line and fields separated by a delimiter, usually a comma. Use it when a person will open the file in Excel or a text editor, when a script appends rows one line at a time, or when the receiving tool only accepts text.

Parquet

Parquet is a binary, column-oriented format that stores the values of each column together, compressed, with a typed schema. Use it to store, query or move large tables when you want smaller files, faster analytical queries and column types that survive between tools.

DuckDB benchmarked both formats in December 2024 on the TPC-H scale factor 20 lineitem table, running each query directly on the file. Parquet is smaller because the values of one column compress well together. It is faster because a query reads only the columns it needs and skips row groups whose min/max statistics rule them out.

CSV stores only text, so every program that opens it guesses each column's type. That is why tools disagree on whether a column of digits is a number or a string, and whether 03/04/2024 is in March or April.

Parquet writes a typed schema into the file, so numbers, dates, booleans and strings keep their types in every tool that reads it. Text files also carry an encoding question: rows.page reads CSV as UTF-8, as UTF-16 when it has a BOM, and otherwise as Latin-1 with the note "Read as Latin-1 text.", but a byte from 0x80 to 0x9F, such as a Windows-1252 euro sign, stops the file from opening until you re-save it as UTF-8.

No. Parquet is binary, so a text editor shows unreadable bytes, and desktop Excel has no Parquet connector; Power Query's Parquet connector exists only in Power BI and Power Query Online. A CSV opens in both, and you can grep it or paste a few lines into an email.

If you only have a browser, drop the .parquet or .pq file on rows.page to sort, filter, check per-column stats and run SQL. DuckDB-WASM reads it on your machine, and nothing is uploaded unless you press Share.

• A script or logger appends records one line at a time as events happen.
• Teammates need to open the file in desktop Excel.
• You want to read or diff a small file in a text editor.
• A downstream tool only accepts text, such as comma-delimited UTF-8.

Parquet is written in row groups with its metadata at the end of the file, so you can't add a single line the way you can with CSV.

1. Open rows.page and drop a .csv, .txt, .parquet or .pq file on the page, or click Paste data to paste cells copied from Excel or Google Sheets.
2. Optionally narrow it down with a filter such as country:NL or price:>20, or a SQL query against the table data.
3. Open Export and pick Parquet for a zstd-compressed .parquet file, or CSV for a comma-delimited UTF-8 file with a header row and no BOM.

Export writes the current view: filter, sort, hidden columns or SQL result. To drop duplicate rows first, exclude the hidden __rid row-number column:

SELECT DISTINCT * EXCLUDE (__rid) FROM data

• SQL-result exports stop at 1,000,000 rows; filtered views export every matching row.
• CSV export has no delimiter, encoding or BOM option.
• Cells that start with =, +, -, @, a tab or a carriage return get a leading apostrophe, so spreadsheets don't run them as formulas.
• rows.page does not read or write .xlsx or .xls, so save Excel files as CSV first.