[hvidgaard]

While i like the novel approach to see if it's possible, I miss some discussion that relates them to other algorithms for known scenarios. Sorting is embarrassingly easy to parallelize. Basically you can Divide'n'Conquer with as many machines as needed with only the networking and storage interface as the limiting factor. Could it be cheaper when using a GPU and ML? Probably, but when the running time is unpredictable and highly dependent on the training and you cannot even guarantee correctness, is it really worth it?

I'd be more interested in seeing the analysis in terms of IO operations; because the CPU will be plentiful when the data set grows beyond what is possible to juggle in the memory.