I like @jwdebelius’ idea of using xarray though admittedly I’m more familiar with it than the others mentioned. One of my thoughts is about dimension order being a concern here - for 2D data structures it’s easy and often intuitive to enforce a standard (sample x feature for FeatureTable, feature x covariate for FeatureData[Differential], etc.) but when we start introducing more I imagine this could cause some frustration when saving and loading FeatureTensor objects.
At the abstracted level I could foresee a headache figuring out which dimension is Feature in some N-dimensional tensor if not named like in xarray. Named dimensions could alleviate some of these issues.