dsdiff

A git-style diff between two dataset files: schema changes and column-level distribution drift, with a CI gate.

When a dataset is regenerated, columns quietly get renamed, retyped, gain nulls, or shift distribution, and the pipeline keeps running while the model degrades. There is no quick "git diff for data" a reviewer can read on a pull request. dsdiff is that: point it at two files and it reports what changed, ranked by how much it should worry you.

$ dsdiff diff yesterday.parquet today.parquet
severity  column      change         detail
high      income      drift          PSI 0.412
high      signup_date column_added   new column
medium    age          null_rate      null rate 0.0% -> 7.3%
low       country      cardinality    distinct values 41 -> 44

Install

$ pip install dsdiff                 # from PyPI, once released
$ pip install git+https://github.com/jmweb-org/dsdiff   # latest, available now

Reads CSV, Parquet and JSON Lines through polars. No services, no schema files to author.

Usage

$ dsdiff diff a.csv b.csv            # human-readable table
$ dsdiff diff a.csv b.csv --json     # machine-readable findings
$ dsdiff diff a.csv b.csv --markdown # a table to paste into a PR
$ dsdiff diff a.csv b.csv --check    # exit non-zero on a high-severity change

Commit a baseline

Profile a dataset once and compare future data against the saved profile, without re-reading the original file:

$ dsdiff profile reference.parquet -o baseline.json
$ dsdiff diff baseline.json new_batch.parquet --check

The baseline stores the bin edges, so drift on new_batch is measured against exactly the same buckets as the reference.

In CI

- run: dsdiff diff baseline.json data/current.parquet --check

What it checks

• Schema: columns added, removed, or retyped (all high severity).
• Nulls: a jump in the null rate of a shared column.
• Cardinality: a categorical column gaining or losing distinct values.
• Distribution drift: the population stability index (PSI) per column, numeric columns binned by quantiles and categoricals by frequency.

Severity and the PSI scale

PSI is the standard tabular drift measure. The conventional reading is used here: below 0.1 is low (no meaningful shift), 0.1 to 0.25 is medium, and 0.25 or above is high. Schema and type changes are always high. By default --check fails only on high-severity findings; pass --fail-on medium to tighten the gate.

Exit codes

License

MIT. See LICENSE.