I was also surprised this used leveldb underneath, especially since the goal is for larger than memory datasets.

For every level you add an additional disk read to every query. For optimal performance you should have RAM greater than 10% the size of your dataset, so the OS disk cache can operate effectively.