--drift-report

Compare two versions of the same dataset and report how it moved.

octa --drift-report yesterday.parquet today.parquet
octa --drift-report a.csv b.csv --fail-on null_rate:0.05,rows:0.1
octa --drift-report a.csv b.csv -f json

Columns are matched by name. One present on a single side is reported as added or removed and produces no metric rows. Every shared column gets its null rate, distinct count and, when both sides are numeric, its minimum, maximum and mean compared. Columns with few enough distinct values also list the category values that appeared and vanished.

The report goes to stdout in the chosen -f format; the row counts, the added and removed column names and the pass/fail summary go to stderr, so a pipe stays parseable.

Output

Column	Meaning
column	Source column, empty for the table-wide rows row
metric	rows, null_rate, distinct_count, min, max, mean, new_values, vanished_values
before	The earlier value, or the vanished category list
after	The later value, or the new category list
change	Relative change, when both sides are numbers
breached	Whether this row failed a --fail-on gate

Gating a pipeline

octa --drift-report baseline.parquet daily.parquet --fail-on null_rate:0.05

--fail-on takes a comma-separated list of metric:change pairs, where change is the largest relative move that still passes. The command exits 1 when any gate is breached and 0 otherwise, so a CI step can depend on it directly. Without --fail-on the report is informational and always exits 0.

A baseline of zero that moved at all counts as an unbounded change. That is deliberate: a null rate going from 0 to 0.5 is exactly the case a gate exists to catch, and a relative change is undefined there.

A gate naming a metric that never appears in the report is not an error and does not fail the run.

Every figure comes from the same Summary pass the Analyse -> Summary tab uses, so the command line and the window can never disagree. See Data Drift for the dialog.