data_drift

Compare two versions of the same dataset and report how it moved: columns added or removed, and per shared column the change in null rate, distinct count and, for numeric columns, minimum, maximum and mean. Columns with few distinct values also report which category values appeared and vanished.

Read-only, so it stays available under --mcp-read-only.

Parameters

Name	Type	Meaning
path_a / path_b	string	The earlier and later versions. May be cloud URLs.
open_tab_a / open_tab_b	string	Use an open GUI tab instead (name, or @active).
table_a / table_b	string	Sheet or table name for multi-table sources.
category_cap	number	Skip category comparison above this many distinct values. Default 50.
fail_on	string	metric:change gates, e.g. null_rate:0.05,rows:0.1.
unlimited	bool	Lift the streaming row cap for this call.

Response

{
  "rows_before": 1000,
  "rows_after": 1000,
  "added_columns": ["region"],
  "removed_columns": [],
  "failed": true,
  "drift": [
    {
      "column": "amount",
      "metric": "null_rate",
      "before": 0.0,
      "after": 0.5,
      "before_text": "0",
      "after_text": "0.5",
      "change": null,
      "breached": true
    }
  ]
}

failed is true when any gate was breached, so it can be used directly as a pass/fail. A baseline of zero that moved at all counts as an unbounded change, which is why change can be null on a breached row.

This measures distributions, not rows. Use diff_tables when the question is which rows changed.

See Data Drift.