Currently there is a 1-1 mapping between TileDB dimension as index columns in SOMA objects and TileDB attributes and non-index columns. The datatype for index columns should match the supported types for TileDB dimensions and the same holds true for attributes as well.

With the introduction of the GeometryDataFrame the two previous condition are not true anymore. soma_geometry is an index column with binary datatype holding WKB encoded geometries. TileDB doesn't support binary blobs as dimensions out of the box and even if it did, spatial indexing would require decoding the underlying WKB blob. To provide spatial indexing GeometryDataFrame utilizes addition dimensions and makes the actual soma_geometry an attribute of type WKB.

Dimensions and attributes generated internally for indexing should be hidden to the end user and any kind of interaction with that column should be through the original index column name.

This approach requires special handling of that column in multiple places throughout the codebase. Any column in the future that would require special indexing will also face the same issues and would increase the complexity of the codebase.

(UPDATE: SOMA specification removed index columns from GeometryDataFrame with the possibility to add it back at a later stage)