derive

derive adds one computed column. It is the right node when a feature deserves a name and should be inspectable in the DAG.

Prefer a chain of small derives over one opaque script when each intermediate feature is useful for review. Use a language node when the computation needs loops, model code, external packages, or multiple output columns at once.

Feature contract

Field	Required	Notes
inputs	yes	Exactly one upstream table.
as	yes	Identifier-shaped output column name.
expr	yes	Expression compiled to a Polars expression and aliased to as.

Writing the expression

• expr is compiled to Polars and aliased to as, so the expression itself does not need an assignment.
• Use coalesce() when nulls should become a default value instead of following native null behavior.
• The new as column cannot collide with an existing column. Drop or rename first if you mean to replace something.

Reviewing the result

The preview should make the new column easy to find. For numeric features, a distribution/profile is usually more useful than a long row sample.

default is the input table plus the new column.

Example

- id: lab_load
  kind: derive
  inputs: [patient_lab]         # length 1
  as: lab_load                   # new column name
  expr: "[crp_mean] * [ldl_max] / 1000.0"

Related

• Expression language - supported operators and functions
• select - narrow or reorder columns after deriving