And so the size of your data structures matters. I'm processing lots of data frames, but each represents a few dozen kilobytes and, in the worst case, a large composite of data might add up to a couple dozen megabytes. It's running on a server with tons processing and memory to spare. I could force my worst case copying scenario in parallel on each core, and our bottleneck would still be the database hits before it all starts.

It's a tradeoff I am more than willing to take, if it means the processing semantics are basically straight out of the textbook with no extra memory-semantic noise. That textbook clarity is very important to my company's business, more than saving the server a couple hundred milliseconds on a 1-second process that does not have the request volume to justify the savings.