ALL-DATA0002 · evidence

This page continues ALL-DATA0002 directly from its source docstring. Read its implementation.

Evidence

Each finding records the source location, the asset identifier, and the field name the schema does not declare. The value is the number of reads naming a field that does not exist.

The repair rewrites the literal that named the retired column so it names the column the catalog itself proves replaced it, and it is offered only when the provider retained that proof. Column-level lineage naming exactly one surviving successor is the whole evidence, so a column with two successors, none, or a name the literal spells more than once is reported without a patch. MCMR reparses the rewritten file and reruns this rule before keeping the edit.

Exceptions

A read of a field that exists is not judged here even when its type disagrees, since that is what ALL-DATA0003 answers. A read whose asset is missing is excluded entirely, so one root cause is never counted twice. A field a query selects through a wildcard or builds by interpolation names nothing exactly and never reaches this stream, which under-reports and never over-reports. One literal receives at most one rewrite in a run, because two edits to the same string would overlap, so a query retiring two columns closes over two verified runs.

References

• Cites “dbt documentation”, catalog artifact schema
• Cites “OpenAPI Specification”, property compatibility principles
• Cites “DataHub documentation”, column-level lineage