Changelog#

0.2.0 — 2026-09-24#

read_csv, parse_csv and to_csv take a keyword-only delimiter: ";", "\t" for TSV, or any other single ASCII character but a quote or a line break.
df[["name", "age"]] and df.select([...]) pick columns into a new frame.
df.sort_values(by, ascending=True): a stable sort by one column, radix for numbers, faster than pandas' stable sort on the benchmark.
from_dict({...}) builds a frame from Python data, typing each column by its values; df.to_dict() is the inverse.
An euspinolia command: stats for types, ranges and distinct counts per column, head for the first rows. Also python -m euspinolia.
Fixed: in a one-column frame, to_csv wrote an empty string as a blank line, which the parser skips, so the row was lost on reading it back.
Column lookup no longer accepts a bool as a position.
The C ABI changed (eus_read_csv, eus_parse_csv and eus_frame_to_csv take a delimiter); the library and its Python layer ship together, and self_check catches a mismatched pair.

Wheels for 32-bit ARM Linux (armv7l), so Raspberry Pi OS installs do not fall back to a source build; Zig 0.16.0 cannot build on a 32-bit host.

First release.

read_csv / parse_csv: a subset of RFC 4180 parsed in Zig into typed columns (int, float, string), with per-column type inference.
DataFrame and Column: shape, names, dtypes, indexing, slicing, iteration, head, row; numeric columns are read through the Zig buffer without copying.
Reductions: sum, mean, min, max, keeping the column's type.
Filtering: df.filter(column, op, value) and df[df["age"] > 30], with & to chain conditions.
groupby(key).agg({...}) with sum, mean, min, max, count, plus the .sum() / .mean() / .min() / .max() / .count() shortcuts.
to_csv: serialise a frame back out, round-tripping types.
Wheels for Linux (x86_64, aarch64), macOS (x86_64, arm64) and Windows (x64, arm64), all cross-compiled from one machine since the library links nothing, not even libc. The sdist builds anywhere with pip, pulling Zig from the ziglang package.