Parquet vs CSV vs JSON: why your 2 GB CSV is a 200 MB Parquet

Your 2 GB CSV is a 200 MB Parquet. Same rows, same columns: the difference is the format tax, and your storage bill and query times both pay it. One page on picking file formats on purpose. The print-ready A4 PDF is at the bottom.

The three

• CSV: rows of text, no types, parse every byte.
• JSON: nested documents, keys repeated on every row.
• Parquet: columnar, typed, compressed. Built for analytics.

Size

df = pd.read_csv("sales_2026.csv")   # 2.1 GB

df.to_parquet("sales_2026.parquet")  # 205 MB

# and reads pick their columns
pd.read_parquet("sales_2026.parquet", columns=["date", "amount"])

Per-column compression is the whole trick: similar values sit together, so dictionary and run-length encoding crush them. JSON is the biggest of the three because the keys ride along on every single row.

Speed

• Columnar reads: load the 3 columns you asked for, skip the other 47.
• Predicate pushdown: row-group statistics let filters skip whole chunks unread.
• CSV scan: parse everything, always, to answer anything.

Types

When each wins

• CSV: humans, Excel, one-off exchange with systems you do not control.
• JSON: APIs, events, genuinely nested data.
• Parquet: lakes, warehouses, anything a query engine scans.

Gotchas

• CSV delimiter roulette: commas inside quotes, encodings, BOMs.
• Parquet is not appendable row by row: write new files, partition by date.
• Parquet is not human-readable: keep a small CSV sample next to it for eyeballing.

Interview phrasing worth memorizing: CSV is a serialization of strings, not of data. Exchange in CSV, run analytics on Parquet.