Parquet vs CSV: Size, Speed and When to Use Each
The one-paragraph answer
CSV is a text interchange format: universal, human-readable, schema-free. Parquet is a storage and analytics format: typed, columnar, compressed. Use CSV to move data between humans and arbitrary tools; use Parquet when the data will be stored, scanned or queried more than once.
Size
Parquet stores each column contiguously and compresses it. Real-world tabular data typically shrinks 5–10x versus the equivalent CSV. That is not a rounding error — it is the difference between an email attachment and a download link, or between a $50 and a $500 monthly S3 bill.
Speed
- Column pruning:
SELECT revenue FROM salesreads only the revenue column from Parquet; with CSV every byte must be parsed. - Predicate pushdown: Parquet row-group statistics let engines skip entire chunks that cannot match a filter.
- No re-parsing: types are stored once, not re-inferred on every read.
Type safety
CSV has no types: 01234 might be a zip code or the number 1234, and every
consumer guesses independently — a classic source of silent data bugs.
Parquet fixes the schema at write time, including timestamps with timezones,
decimals with precision, and nested lists and structs.
Tool support
CSV opens in literally everything, including Excel. Parquet is native to the data stack — Spark, DuckDB, pandas, Polars, BigQuery, Snowflake, Athena — but not to spreadsheets or most SaaS import screens.
Converting between them
Both directions run locally in your browser on this site: CSV to Parquet infers types automatically and writes ZSTD-compressed output, and Parquet to CSV flattens typed columns back to text for spreadsheet users. To sanity-check a file before or after, use the Parquet Viewer.
Frequently asked questions
- How much smaller is Parquet than CSV?
- Typically 5-10x smaller for tabular business data, thanks to columnar layout plus compression like ZSTD or Snappy. Highly repetitive columns compress even further.
- Is CSV ever faster than Parquet?
- For tiny files that are written once and read once, CSV's simplicity can win — there is no metadata overhead. For anything queried repeatedly or selectively, Parquet is faster because engines read only the columns and row groups they need.
- Why does every warehouse still accept CSV?
- Ubiquity. Every tool ever written can produce CSV, so it remains the lowest common denominator for data exchange — which is exactly why converting at the boundary is so common.