This paper targets datasets of 1-10 billion keys, keeping write amplification down to at most 1x per level (i.e. 5x over 5 levels), and in-memory indexes to a minimum.

Here are the slides: https://www.usenix.org/sites/default/files/conference/protec...