Great paper. Enjoyed seeing the RAMCloud callout, as it was the first thing that popped into my head when reading it.

I've been thinking about a few other potential variants. Primarily the 3AZ scenario where you'd use Btrlog in 1AZ, and then send a replica to an S3-Express bucket in another AZ synchronously (a PUT to the local AZ endpoint does not incur a cross AZ data transfer cost, even if the bucket is in a different AZ), and then an async PUT to s3-std.