I only have a cursory knowledge on LMDB (listening to a podcast while biking). Anyway, LMDB has no transaction log nor write ahead log. There's no overwrite during update. Data page update is copy-on-write and b+tree index update is append only. The update on the b+tree pages is performed from the bottom of the tree to the root, linking newly appended pages to higher level pages. The transaction is committed when the new root page is appended. When there's a crash, the incomplete appended index pages have not been linked up to the root page yet and are not reachable from the previous valid root page. They can be just thrown away. Recovery just means searching for the last valid root index page. There's no need for a WAL and undo/redo of the transaction log.

Deleted pages and obsolete pages are actively put back into a free list (tracked by another b+tree), which will be reused for new page allocation. This avoids the long garbage collection phase to walk all the live pages for compaction (no vacuum is needed).