filter

A filter node is a named row gate. The schema stays the same; only the set of rows changes.

Good filter nodes read like cohort decisions: adults only, visits after baseline, active accounts, non-null outcomes. If the expression needs a paragraph to explain it, split the logic into an upstream derive with a readable feature name.

Filter shape

Field	Required	Notes
inputs	yes	Exactly one upstream table.
expr	yes	Boolean expression evaluated per row. Truthy rows are kept.
metadata.label	no	Use a readable label such as “Keep visits after baseline”; the expression itself is usually too terse for reviewers.

Expression guidance

• Write a boolean expression such as [age] >= 18 or [status] == "active".
• Use bracketed column refs and plain literals. Row-level functions like coalesce([score], 0) are fine.
• Do not hide aggregations inside a filter. Build summaries with aggregate, then filter the summarized table.

Reviewing the result

The important review question is row loss. Compare input rows to output rows and make sure a zero-row result is intentional.

Expression parse or evaluation errors fail the node and downstream dependents. The best UI and report copy should point at the expression, not the whole DAG.

Example

- id: adults
  kind: filter
  inputs: [patients]            # length 1
  expr: "[age] >= 18"

Related

• Expression language - syntax for row predicates