Thanks for this. I really feel like I understand the algorithm, and more importantly, the tradeoffs involved, after reading that.

My company makes a product that involves a sort of primary feature that's insanely read heavy (something like 10k to 1 read to write ratio), and a secondary feature that's exactly the opposite. We just use Postgres with for both (and it's obviously the winner for the read heavy workload) but I think I'll have to look more into this stuff.

One interesting thing to note about our write heavy workload is that the reads are usually in chronological order, with a filter. Anyone know what that implies for LSM based stores?