Container classes definitely have their advantages and some systems make great use of them. vtk-m basically uses this approach in how they access raw data through flexible APIs which can handle all sorts of cases. I do think they work best in cases such as vtk-m where the library has been designed that way from the start. VTK has gone with an approach where the choice of data class (ImageData) heavily defines and constrains the API and data. In that context, adding an incremental feature to the existing ImageData is probably going to be easier (maybe not better) than a broader re-architecture of the data model which can be tricky.