I worked on a commercial product that did just this, except it also had a "transaction log" component. All access through the server was logged, and the log could be used for replay/recovery. While it worked for them for thousands of simultaneous clients, it was not a general purpose solution. There were no real indexes: the "indexes" (hash tables, really) were built in memory at startup.
I suppose I should've said "persistent" indexes. There was also only a single type supported: hash table. And it was hard coded to 2 fields. You had to change the code (which was C) to change the database.
This system was rarely restarted, so in practice it didn't matter what it did at startup, as long as it didn't take more than a few minutes, but it did place some limitations on data size. (This was a 32-bit system and everything was memory mapped.)
