Patching LMDB: how we made Meilisearch’s vector store 3× faster

This post details a significant optimization to Meilisearch's vector store by patching LMDB to enable concurrent read transactions on uncommitted write transactions. This was achieved by implementing nested read-only transactions within a parent write transaction, allowing multiple threads to read uncommitted changes lazily on demand. This eliminates the need for a large, slow-to-initialize HashMap to store pointers, reduces cache pollution from full scans, and enables indexation to scale linearly with the number of CPUs, improving time complexity.