Anybody who has ever taken a numerical analysis course understands that
floating-point arithmetic on computers is a messy affair. Even so, it is
easy to underestimate just how messy things can be. This topic came to the
fore in an initially unrelated python-ideas mailing-list thread; what
should the Python statistics
module do with floating-point values that are explicitly not numbers?

Source: LWN.net – [$] Some median Python NaNsense