LSM-tree also covers the goal of maintaining just 2 sub-trees and a concurrent process of merging sub-trees. That sounds like it would take a lot of additional time to get right and would need some off-line process to perform the merge.
Not necessarily.
Merging means applying insertions/deletions from one subtree to another. While it's normally preferable and more efficient to do it in batches, I've successfully implemented in-memory versions that use other writers to perform the task, amortizing the cost of merging across many operations. In essence, when there's a need to merge two subtrees, an inserting process also merges one entry, so slowly trees get merged. That works in-memory very well, it's quite clear that it's not necessarily generalizable to external storage, but it's a technique to have in mind.
Alternatively, vacuum could do it. It's quite clearly a vacuuming task anyway.
