Large and Small are so contextual. I'm processing 350m events/day in 24h splits, and I managed to stop worrying about locality of reference because I'm the sole

Large and Small are so contextual. I'm processing 350m events/day in 24h splits, and I managed to stop worrying about locality of reference because I'm the sole occupant of the machine. When I did worry about it, I found radix tree, awk hash and perl/python hash/dict pretty much all occupied much the same space and time but a tuned C implementation got 2-3x faster than any of them. Somebody else pointed out memory resident for most of this would be faster still but you have to then work to process 24h of things against a single memory instance. Which means buying into IPC to get the data "into" that memory. It interested me you didn't show the idea in rust. That was the only point I was making: Python as pseudocode to think things in documents is fine with me. But overall, I liked your outcome. I just think it's important to remember large and small are very contextual. Your large case looks to me to be >5m things and for an awful lot of people doing stupid things, 5m is bigger than they'll ever see. If the target was only people who routinely deal in hundreds of millions of things, then sure.