Core Concepts
Dataset Diff Engine
Detecting schema drift and quality regressions across dataset snapshots
In production ML environments, datasets evolve continuously as new data batches arrive daily or weekly. Upstream pipeline updates, database migrations, or third-party vendor changes can introduce silent regressions into fresh snapshots.
What Dataset Diff Compares
Featuresmith's fs.diff(old, new) engine profiles both snapshots and computes deterministic deltas across:
- Schema Changes: Added, removed, or renamed columns, and data type changes.
- Structure Changes: Row count deltas and column count shifts.
- Missing Value Spikes: Per-column missingness shifts classified as new, resolved, regressed, or improved.
- Duplicate Shifts: Changes in duplicate row counts and percentages.
- Constant Column Changes: Newly constant or no longer constant columns.
- Cardinality & Statistic Deltas: Shifts in unique values, mean, median, min, max, and standard deviation.
- Leakage Status Deltas: Target leakage findings that were added, removed, escalated, or de-escalated.
Health Verdicts
unchanged: No material structural or quality changes between snapshots.improved: Quality metrics improved (e.g. missingness decreased, leakage eliminated).regressed: Quality degraded (e.g. columns dropped, missingness spiked, schema broke).