Sample Feather File

Feather (v2) is the Apache Arrow IPC file format: Arrow’s in-memory columnar layout written to disk essentially verbatim, with a schema block up front and record batches behind it. Files carry .feather, .arrow, or .ipc extensions interchangeably — they are the same format.

Because the on-disk bytes match Arrow’s in-memory representation, reading Feather is close to a no-op: a reader can memory-map the file and use the buffers directly, with zero parsing and zero copying. That makes it the fastest way to move a dataframe between processes and languages — pandas to polars, Python to R — when both ends speak Arrow. The embedded schema travels with the data, so salary arrives as an int32 column with no type inference, same as the Parquet sample.

This sample is deliberately uncompressed. Feather supports optional LZ4 and ZSTD compression of record batches, and many writers enable it by default — but browser builds of the apache-arrow JavaScript library ship without those codecs, so compressed files fail in web tools with a codec error. Writing with compression='uncompressed' keeps the file readable by every Arrow implementation, including the site’s own Feather to CSV converter. At 30 rows the size cost is irrelevant; at scale, you’d enable ZSTD for storage and accept the narrower reader support.

Feather versus Parquet is a recurring question, and the honest answer is: Parquet for storage (heavier encoding and compression, broad ecosystem, predicate pushdown), Feather for interchange (minimal encode/decode cost, exact Arrow fidelity). If you need this data as Parquet, convert via CSV to Parquet or inspect it in the Parquet Viewer.

All eleven format samples in this section carry the same 30 employee rows, so converted output can be checked against the CSV baseline.