NumPy arrays can be persistently saved on disk using built-in functions in NumPy such as np.savetxt, np.save, or np.savez, and loaded in memory using analogous functions. These methods are best when the arrays contain less than a few million points. For larger arrays, these methods suffer from two major problems: they become too slow, and they require the arrays to be fully loaded in memory. Arrays containing billions of points can be too big to fit in system memory, and alternative methods are required.

These alternative methods rely on memory mapping: the array resides on the hard drive, and chunks of the array are selectively loaded in memory as soon as the CPU needs them. This technique is memory-efficient, at the expense of a slight overhead due to hard drive access. Cache mechanisms and optimizations can mitigate this issue.

The previous recipe showed a basic memory mapping technique using NumPy. In this recipe, we will use a package named...