We have abstracted the database into a ClientDB trait and implemented a DiskDB using RocksDB for production.

It would be useful to have an implementation of an in-memory database that can be used in testing. The benefits would be speed (i.e., RAM is faster than disk) and convenience (we don't need to touch the filesystem).

The in-memory database doesn't need to be particularly optimized or elegant, just the fact it's in memory should make it fast enough. I would suggest a two-layer nested hashmap to simulate columns. E.g., HashMap<HashMap<Vec<u8>>>. It will need to implement some Mutex/RwLock internally so it can be sent around threads like RocksDB can be.

It should live in a memory_db.rs file here and I think the struct should be called MemoryDB.