Honestly, I didn't know any better? From what I've see with implementations as seen in https://github.com/wspeirs/btree the entire BTree needs to fit in the mmap buffer and than re-written entirely on each read / write which seemed limiting both in terms of flexibility (i.e. using variable sized keys and values using an offset table) and in terms of how much keys each node can take. However I didn't do precise math, I would love to hear if you have any insights, what are the pros and cons of using mmap files vs implementing paging from scratch?