To clarify, as I think some people have misunderstood: we used an existing library called bknr.datastore to handle the "database" part of the in-memory store, so we didn't have to invent too much. Our only innovation here was during the Expand phase, where we put that datastore behind a Raft replication.