I have always found CRDTs a lot easier to wrap my brain around. Agreed with Joseph that optimizing them is non-trivial. In Teletype for Atom, we indexed fragments in splay trees, and each text fragment was actually a member of two splay trees at the same time: one for the full document, another for the pieces of the original insertion as they'd been split apart. We would find the insertion fragment in the insertion tree, then walk parent pointers in the document tree to locate it. In Zed, we have two trees, but can't do this double embedding because we use persistent data structures where parent pointers are not an option. So we instead have insertion trees that map subslices of inserted text to fragment ordered identifiers, which we use for lookup in the main document B-tree. These fragment identifiers were themselves inspired by another CRDT paper called Logoot, but we don't actually send them over the wire.