The problem was that an in memory index, because of its size, did not have the same level of redundancy on a per-cluster basis as a disk-based index (e.g. usually every page was present in the memory of just a single machine).

→ View original post on X — @jeffdean,