Holy cow, using mmap batching to stream chunks.npy input in #Julialang is a difference between 70gb of ram over 7 minutes and essentially nothing over 3 seconds. This really opens doors for different analysis on my machine!