Model runs#

One organization-wide model run evaluates the complete model graph at a fixed data-repo revision. A model starts only after its upstream models have succeeded.

Dependency order#

Supernova parses lake references from each model and merges them with explicit -- depends_on: directives. CTE names are local and never become dependencies. Cycles are refused before any model starts.

The graph page shows the same nodes and edges used by the run. Folder order, filename sorting, and modification time have no effect.

Type checking#

Organization policy can set typed models to off, warn, or enforce. In warn, diagnostics are recorded and the model may continue. In enforce, a refused check or remaining error stops that step.

Each successful check records a type contract digest. A later shape change is marked as contract drift so reviewers can distinguish a changed output from an unrelated runtime failure.

Manual models#

-- cadence: manual keeps a model in the graph but skips it during recurring organization-wide runs. Use Run now when you intend to refresh it. Downstream dependencies still make its role visible.

Materialization#

Without -- unique_key:, the new result replaces the output table. With a unique key, rows upsert by that key. Both modes publish a complete table commit; readers do not observe an intermediate state.

Model statements have a roughly 4-hour reaper. Eligible memory failures may retry with a larger memory budget. The stored run preserves each attempt and the final outcome.