diffolars 1.1.2

A small Polars-based toolkit for comparing two versions of a dataframe and generating randomized test data to exercise that comparison.

diffolars

A small Polars-based toolkit for comparing two versions of a dataframe and generating randomized test data to exercise that comparison.

Ideally used to compare dataloads in the day-to-day of a database analyst.

GitHub: https://github.com/ko222uky/diffolars

Installation

uv add diffolars

Generating test data

diffolars.demo generates a random dataframe and a mutated copy of it, useful for testing diff logic without hand-crafting fixtures.

from diffolars.demo import get_df_pair

pair = get_df_pair(
    n_rows=100,
    n_cols=10,
    n_new_rows=5,    # rows added in the mutated copy
    n_new_cols=2,    # columns added in the mutated copy
    coverage=0.1,    # fraction of existing cells randomly changed
    seed=42,
)

original = pair["original"]
mutated = pair["mutated"]

Every generated row gets a record_id UUID column, used to match rows between the original and mutated dataframes. get_random_data and get_mutated_data are also available individually if you want to generate or mutate a dataframe on its own.

Diffing

diffolars.diff provides the comparison API for two dataloads (e.g. a previous load vs. the latest load), and is under active development.

Inputs to these functions can be a list, polars.DataFrame, or polars.LazyFrame.

• column_intercept / column_symmetric_diff — shared vs. exclusive columns between the two tables
• row_intercept / row_symmetric_diff — shared vs. exclusive rows, based on a record ID column
• prune_rows — returns the rows exclusive to each table (i.e. dropped by the join), tagged with which table they came from
• report_prune — summarizes the pruned rows/columns as a single dict entry, suitable for logging
• get_core — prunes each table down to their shared rows and columns, and returns them as a pair of _A/_B-suffixed DataFrames ready for a field-to-field comparison
• bitdiff — joins the two core tables and computes a per-row diff_bitarray (pl.UInt64), with each bit flagging whether a given column matched between the previous and latest load. Since each row's diff is packed into a single 64-bit integer, the core tables can have at most 64 non-ID columns; compute_bitarray64 (the per-row helper bitdiff calls) raises a ValueError if that limit is exceeded
• bitdiff_summary — reads back a bitdiff result and reports, per core column, how many rows were modified vs. not modified
• bitarray_upset_plot / bitdiff_plot — builds an upset plot (matplotlib) showing which columns tend to be modified together, with an optional top_n to limit the plot to the most-frequently-modified columns and a left-hand histogram of each column's total modification count

Currently only the Windows build is available.

Command-line interface

diffolars.cli exposes diff_cli, a Click command that runs the diff pipeline (report_prune, pruned_rows, bitdiff) over two parquet dataloads and prints the three result tables. It's registered as the diffolars console script.

From within this project:

uv run diffolars \
  --prev-load original.parquet \
  --latest-load mutated.parquet \
  --id-col record_id

Or by using uvx:

uvx diffolars \
  --prev-load original.parquet \
  --latest-load mutated.parquet \
  --id-col record_id

Option	Default	Description
--prev-load	original.parquet	Path to the previous/original data load.
--latest-load	mutated.parquet	Path to the latest/mutated data load.
--id-col	record_id	Name of the record identifier column.
--scan / --no-scan	--scan	Read with pl.scan_parquet (lazy) instead of pl.read_parquet (eager).
--write / --no-write	--write	Write the resulting diff tables to parquet.
--bitarray-summary / --no-bitarray-summary	--bitarray-summary	Produce a per-column modified/not-modified summary and upset plot after the bitdiff is computed.
--top-n	20	Limit the upset plot to the top N most-frequently-modified columns.

When --write is set (the default), results are saved under data/<prev-stem>-<latest-stem>/<today's date>/, as diff_activity_log_record.parquet, diff_record_differences.parquet, and diff_bitarray_results.parquet. If --bitarray-summary is also set, this directory additionally gets bitarray_summary.parquet and bitarray_summary_upsetplot.png.

API reference

Static API docs are generated with pdoc from the package's docstrings, and are hosted at ko222uky.github.io/diffolars (source lives under docs). Regenerate them after docstring changes with:

uv run --group dev pdoc diffolars -o docs

Testing

Tests performed with pytest and coverage:

 uv run coverage run -m pytest --junitxml=report.xml ;
 uv run coverage html

Test result data is created in data/prev-latest.

License

Dual-licensed under either of

• MIT license (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)

at your option.