csvcomp

Semantic CSV↔CSV diff (schema + rows + cells) with zero runtime dependencies.

csvcomp compares two CSV/TSV files and reports:

• Schema changes — columns added, removed, or (optionally) renamed.
• Row changes — rows inserted, deleted, or modified. Rows can be matched positionally or by a key column (or composite key).
• Cell changes — for each modified row, exactly which cells changed.

Everything is zero runtime deps — dependencies = [] in pyproject.toml. The whole library runs on Python 3.11+ stdlib only (csv, io, datetime, decimal, collections, difflib, json, argparse).

Why this exists

Existing CSV-diff tools either refuse the most common production case, or aren't embeddable in Python:

• pandas.DataFrame.compare() raises ValueError on the first shape mismatch — a single added column in a new export breaks the diff. (Quoted verbatim from the pandas docs.)
• paulfitz/daff is a Java library with a Node CLI wrapper; no embeddable Python API.
• csvkit ships 13 subcommands; none of them is a diff.
• yourlabs/csv-diff targets JSON-line key-based diffs, not CSV↔CSV.

csvcomp fills the gap: a single-file-library pure-Python diff that works in CI, scripts, and pipelines without dragging in pandas, numpy, or any third-party CSV parser.

Install

The package is not yet on PyPI. From a clean clone:

pip install git+https://github.com/prasad-a-abhishek/csvcomp.git

For local development:

git clone https://github.com/prasad-a-abhishek/csvcomp.git
cd csvcomp
pip install -e ".[test]"
pytest

CLI usage

csvcomp a.csv b.csv --key order_id --coerce amount:number,shipped_at:date --format unified

@@ columns @@
  order_id   (kept)
  amount     (kept)
  shipped_at (kept)
+ tracking_id (added)

@@ rows @@
  key=[1042]  ~row  amount '199.99' -> '219.99'
  key=[1099]  +row  order_id='1099'  amount='10.00'  shipped_at='2026-01-15'
  key=[1101]  -row  order_id='1101'  amount='5.00'   shipped_at='2026-01-12'

Exit codes

Common flags

Python API

from csvcomp import diff

result = diff(
    "a.csv",
    "b.csv",
    keys=["order_id"],
    coerce={"amount": "number", "shipped_at": "datetime"},
    detect_renames=True,
)

result.added_columns      # list[str]
result.removed_columns    # list[str]
result.renamed_columns    # dict[old, new]
result.added_rows         # list[dict]
result.removed_rows       # list[dict]
result.modified_rows      # list[RowChange]
result.changed_cells      # list[CellChange]
result.is_empty           # bool — True if no changes at all
result.format("json")     # str — JSON representation
result.format("unified")  # str — git-style diff
result.format("sidebyside")  # str — two-column table

Coercion kinds

Unknown coerce kinds raise ValueError (subclass UnknownCoercionError — also a CsvCompError).

NA semantics

The following cell values are treated as equal to "" for comparison purposes (case-insensitive, surrounding whitespace stripped):

"", na, n/a, null, none, nan, -

This means two rows that differ only in which NA spelling they use are considered identical.

Limitations / non-goals

These are deliberately out of scope, per spec:

• No JSON / XML / Parquet / Excel support. CSV/TSV only.
• No streaming / chunked diff. Operates on in-memory dicts. Files in the ~10k row range are fast; files in the millions are not.
• No three-way diff. Two files at a time.
• No patch-file output. Read-only diff.
• No HTML / ANSI-coloured output. Three formats only: JSON, unified text, side-by-side table.
• No pandas adapter. Zero-runtime-deps invariant forbids it.
• No cursor / TUI.
• Rename detection has a 0.80 left-coverage threshold and is OFF by default. Renames of columns whose values also legitimately appear in another column (e.g. country_code vs country_iso2) may mis-detect. See src/csvcomp/diff.py for the algorithm.
• Encoding is UTF-8 only (with BOM stripping). Other encodings raise ParseError.
• Dialect sniffing is unreliable on samples <8 KB. The spec calls this out (KeitaNakamura/diff-csv issue #1). Use --dialect explicitly for production data.
• Trailing fully-empty rows are silently dropped. This is intentional (matches pandas.read_csv's default skip_blank_lines=True) and is documented in the spec.

Development

pip install -e ".[test]"
pytest -v

The test suite has 126 tests covering every spec acceptance criterion. See tests/COVERAGE.md for the criterion-to-test mapping.

License

MIT. See LICENSE.