It feels to me like this just a kludge added to deal with a lack of a stateful iterator on top of the tree. In this use case, the author indicates that it is about exploiting locality in a request path. Imagine that the tree offered an iterator that maintained a stack of its path and that thing had a seek method. You can optimize that thing to only go back up to the root if it needs to. Stateful iterators make for a nice pairing with btrees. These other hints are papering over the lack of that abstraction as far as I can tell.