Compare two Parquet files in the browser
Load File A and File B as local .parquet files. After both files are ready, the tool automatically reports schema differences, row counts, row-group counts, and record comparison results. Field-level changes apply only to common flat columns.
Choose a comparison mode
- Keyed (match by key column(s)) matches records using selected common flat key columns and can report added, removed, and changed records.
- Row order (match by position) compares the record at each position in File A with the record at the same position in File B.
- Row-multiset hash (ignore order, count-only) ignores order and identity, reporting count-based added, removed, and common results without field-level values.
How to compare the files
- Load both local files in File A and File B using the drop areas or browse files.
- Choose a value in Record comparison mode.
- For Keyed (match by key column(s)), select common flat columns under Key column(s) - required for keyed mode.
- Wait for the automatic comparison to finish after both files and the comparison settings are ready.
- Review Row groups & row counts, Schema comparison, and Record comparison.
Understand keyed matching and duplicate keys
Choose key columns that uniquely identify records. Duplicate key values are matched by order of appearance, and the tool displays a warning. Results can therefore be ambiguous when the selected key is not unique. The selected key columns are excluded from changed-field comparison.
If the files have no common flat columns, keyed comparison is unavailable. Use Row order (match by position) or Row-multiset hash (ignore order, count-only) when appropriate.
What the results and exports contain
Schema comparison reports field names and available physical/logical type, repetition, or nullability details. It can list fields added in File B, removed from File B, or changed in both files. Nested or group columns remain in the schema but are excluded from field-level record comparison.
Record results can include added, removed, and changed records, depending on the selected mode. The visible added, removed, and changed tables show at most the first 500 records. Added and removed tables show at most 10 columns.
Export JSON includes schema and record results, but not the displayed row-group statistics. Export CSV contains record-level statuses or row-multiset summary counts; it does not contain the schema diff or displayed row-group statistics.
Limitations and local processing
The reader processes Parquet data row group by row group in the browser. A complete comparison can still require substantial browser memory. If the reader cannot decode a file's type or encoding, the tool reports an error instead of completing an incomplete comparison.
The selected files are passed to the browser-side reader for comparison. This workflow does not upload or modify the source files. The implementation does not define a complete supported-type list, browser compatibility requirements, file-size ceiling, account policy, or formal usage-limit policy.
The tool does not display or compare column compression/statistics metadata or per-row-group metadata.
Frequently asked questions
Why not just convert both files to CSV and diff those?
Converting to CSV changes the data into flat text and does not preserve the Parquet schema details shown by this tool, such as field names, reported physical/logical type details, repetition or nullability where available, and row-group counts. This tool reads the selected Parquet files directly in the browser.
What's the difference between keyed, row-order, and row-multiset hash comparison?
Keyed (match by key column(s)) uses selected common key columns. Duplicate keys are matched by order of appearance and produce a warning. Row order (match by position) compares records by position. Row-multiset hash (ignore order, count-only) ignores order and identity and reports count-based results without field-level values.
Why does keyed mode require a key column?
Keyed (match by key column(s)) needs one or more common flat columns to match records. If duplicate values occur, matching follows order of appearance, so the selected key should be unique. Keyed comparison is unavailable when there are no common flat columns.
How does this handle large Parquet files?
The reader processes data row group by row group in the browser. A full comparison can still use substantial browser memory.
Does this tool upload my files anywhere?
The selected files are passed to the browser-side reader. The comparison workflow does not upload or modify the source files.