What the File Knows

A query filters one day out of a year of Parquet files and reads 2 of 80 columns. It ends up reading a small fraction of the bytes on disk even though there is no external index. Which property of the format allows that?

Parquet files are always physically sorted by ingestion timestamp, so a reader binary searches to the day boundary and stops once it passes the end of the range

Each file is divided into row groups whose footer records per-column min and max statistics, so a reader skips non-matching row groups and fetches only the needed column chunks

Parquet maintains a global B-tree index across every file in a directory, and the reader consults that index to locate the matching rows before it opens any of the data files themselves

Parquet stores data row by row but compresses each row on its own, so the columns a query never mentions decompress to zero bytes and cost the reader nothing at all