[ARROW-11001] [C++][Dataset] Enable column renaming (in physical schema -> dataset schema) in Dataset scanning

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: C++
• Labels: dataset
• External issue URL: https://github.com/apache/arrow/issues/26921
• Assignee: Unassigned
• Reporter: Joris Van den Bossche
• Created: 21/Dec/20 17:00
• Updated: 11/Jan/23 08:16

Description

Currently, we allow dropping/adding columns when scanning the actual sources of a Dataset (e.g. if newer files in the dataset have additional columns), but we should also provide a way to specify fields that are renamed in certain files of the dataset.

While it might be possible to also provide some convenience for this in the discovery factories, it's probably best to start to see how this could be added to the actual Dataset class and the lower-level constructor functionalities (such as FileSystemDataset main constructor from fragments or from_paths).

What I am thinking right now, is that we would need an (optional) mapping of "field ref/name in physical schema -> name in projected/dataset schema" for each fragment of a dataset.
However, that might not fully fit in the current design, as the fragment doesn't know about the dataset schema, but only sees this when it is projected.

cc bkietz

Issue Links

• is related to: ARROW-11003 [C++][Dataset] Schema evolution in Dataset scanning