I'm not sure why the author is dismissing space-filling curves so quickly, just based on the age of the idea. I've worked with them extensively, in the context of spatial indexing in a database system, in both research and product settings. They address a number of problems described for interval data types. The techniques I've been working with "decompose" a spatial object into a number of regions by recursively partitioning the space, ending up with a small number of "z-values" for each. A z-value is represented by an integer, which is then used as a key in a conventional index, e.g. a sorted array, balanced binary tree, B-tree, skiplist, etc.
Getting a uniform distribution for sharding is easy in one dimension. I'm not sure what kind of scaling the author has in mind, but if you put the z-values in a B-tree, it scales exactly like a B-tree, i.e., extremely well. I agree there is no exploitable ordering of the spatial objects, but the trick is to order the z-values. If a given spatial object gives rise to 4 z-values, then they -- and the spatial object -- appear in 4 places in the ordering. That's fine. Run your search (in one dimension), get qualifying records, and then get rid of duplicates later.
Getting a uniform distribution for sharding is easy in one dimension. I'm not sure what kind of scaling the author has in mind, but if you put the z-values in a B-tree, it scales exactly like a B-tree, i.e., extremely well. I agree there is no exploitable ordering of the spatial objects, but the trick is to order the z-values. If a given spatial object gives rise to 4 z-values, then they -- and the spatial object -- appear in 4 places in the ordering. That's fine. Run your search (in one dimension), get qualifying records, and then get rid of duplicates later.