Data & Analytics
A dbt Fusion Python UDF may deploy more code than you wrote
Inspect the deployed BigQuery routine body during dbt Fusion migration and compare it with the intended function source.
November 8, 2026
A Python UDF may pass dbt build yet contain code that was never in its source file. An open dbt issue reports that Fusion 2.0.9 deployed a BigQuery Python UDF from functions/ with extra dbt scaffolding and build-environment identifiers in the routine body. The reporter compared it with dbt-core, which emitted only the user function. This is a deployment artifact discrepancy, not evidence that every Python model or warehouse behaves the same way.
Compare source with the warehouse object
List the UDF source and YAML configuration under functions/, then build only that function in a non-production project. Read routine_definition from BigQuery INFORMATION_SCHEMA.ROUTINES, as the issue reproduction does. Save a normalized diff between the intended source and deployed body. Record the dbt engine, adapter, package versions, project and dataset. The dbt UDF guide says dbt build creates the warehouse function from the source and property file; the deployed object is the contract downstream SQL actually calls.
Assess what the extra body changes
Determine whether the appended code affects runtime behavior, routine size, reviewability, or visibility of build-time identifiers. Avoid calling the identifiers a security breach without reviewing the routine’s grants and data classification. Run a small deterministic call and compare output across the two engines, but keep the body diff as a separate acceptance check; functional equality does not make an unexpected deployment artifact acceptable.
Gate the migration
Add a routine-definition check to the migration checklist for each language and adapter combination you use. For this reported version pair, hold production rollout of affected UDFs or obtain a tested workaround from dbt. If a later release changes the output, verify the deployed routine again rather than relying only on release-note language. Keep the routine definition and build logs with the support case.
Adovo’s view
The compiled artifact is the product, not the file in the repository. Data teams should review what the warehouse received, especially when an engine migration can alter executable code. A passing build is an input to trust, never the final evidence for it.