For example, consider a program that reads in a git fast-import stream. The first commit creates a file named README.txt. How many more commits will touch the same path? You’ve no idea, of course, but it could be hundreds of thousands. Do you want hundreds of thousands of individual strings all containing that same filename? That would be fine in the first draft of your program, but in the long term it is too big of a cost to pay (believe me, I know). So you want some kind of pool to store filenames. It could be a grow-only arena, so that you can use plain references, or a mutable pool using reference counting (Arc obviously).

Rust gives you the power to start writing your program using ordinary strings for filenames (or OsStr, obviously), and then later graduate to something both more complicated and more efficient just by swapping out the implementation of one type.