WRANG

01Overview

WRANG is a data-wrangling toolkit with three ways in: an interactive, menu-driven terminal session; non-interactive flags for scripts and CI; and a Python library whose modules can be imported on their own. It reads and writes CSV, Excel, Parquet and JSON or JSON Lines.

02Where it came from

WRANG succeeds RIDE-CLI, the command-line half of my master’s project. Installing it adds both a wrang and a ride command, identical aliases, so existing ride scripts keep working.

03Architecture

• Core modules: loader, inspector, explorer, cleaner, transformer and validator, each built on Polars DataFrames, with lazy scanning for large files.
• CLI layer: an interface, menus and formatters that render tables and panels with Rich.
• Optional extras: wrang[viz] adds matplotlib and seaborn, and wrang[advanced] adds DuckDB and ConnectorX.
• Configuration: user preferences such as the outlier factor and chunk size are kept in ~/.wrang/config.json.

04Implementation

• Chainable cleaning and transforms. DataCleaner and DataTransformer return themselves, so a pipeline reads top to bottom. There are nine imputation strategies, including distribution sampling and KNN.
• SQL on any file. --sql runs DuckDB against the loaded file as a table named data.
• Built for automation. --inspect --output-format json prints a machine-readable overview. --chunk-size streams large files, and --compare diffs two datasets.
• Data contracts. A schema can be inferred from data, saved as JSON or YAML, and validated later, with violations reported by severity.
• Reports. --profile writes a self-contained HTML profile of a dataset.

05In use

Real output from wrang 0.2.2, installed from PyPI and run on the Titanic dataset bundled with the repository:

$ wrang titanic.csv --sql "SELECT Pclass, COUNT(*) AS n,
    ROUND(AVG(Survived),3) AS survival_rate
    FROM data GROUP BY Pclass ORDER BY Pclass"

  Pclass   n     survival_rate
 ──────────────────────────────
  1        216   0.63
  2        184   0.473
  3        491   0.242

3 row(s)

The repository’s README reports a test suite of 217 passing tests (plus 2 expected failures), and releases are published to PyPI from GitHub Actions.