Nice write-up on squeezing performance out of Go on the billion-row problem that was making the rounds earlier in the year.

I honestly like r8 (the clear version parallelised) most, I think. Some of the other tricks, like using ints not floats are good (and more accurate anyway), but the later solutions get too optimised for my taste.

(It is true that this could matter in some cases, but I suspect that last 6.5x speed up generally isn’t needed and gets outweighed by the much trickier maintenance).