Skip to main content

Map Version To Schema

POST 
/datasets/:dataset_id/versions/:version_id/map

Apply column mapping and advance stage to mapped.

Reads the version's Parquet (using original_minio_path when available so re-mapping always starts from the original file), applies the column rename specified by body.column_mapping (GDI field → dataset column), and writes the result to a new object, repointing minio_path at it. The write never lands back on an object minio_path/original_minio_path currently name — not the upload, not a version's own former golden snapshot if it was demoted, and not an object shared by pointer copy with a different version's row (dimension-draft minting sets a new version's minio_path to an existing version's current path with no object write). Enumerating which existing objects are safe to overwrite in place has twice missed a case that corrupted a golden dataset, so nothing is overwritten in place at all. Stores the mapping in dataset_metadata["column_mapping"] and sets stage="mapped".

Calling this endpoint a second time with the mapping the version already records is accepted, so retries are safe. A different second mapping re-derives from the original file for versions carrying original_minio_path — that object is never touched, so it is always available to re-derive from. Augmented, trace-landed and ground-truth versions have no upload artifact to re-derive from, so a changed mapping there resolves against whatever the previous mapping produced and is rejected when that no longer has the columns requested.

Rejects with 422 when a mapping value names a column that is not there, which is what an inverted {dataset column → GDI field} mapping looks like; the error lists the columns actually available. Also rejects a mapping that points two GDI fields at one source column, which would otherwise drop one of the two renames. Also rejects with 422 when the version's upload was already overwritten by a historical mapping bug — its original columns cannot be recovered, and re-mapping is refused rather than silently building on the corrupted data; upload a fresh version instead.

Request​

Responses​

Successful Response