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.
Since the WAL itself may contain sensitive application data, would measures such as encryption at rest/in transit, KMS/key management, and access control alter the latency profile observed in current benchmarks?
3 comments
[ 7.6 ms ] story [ 18.0 ms ] threadGreat 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.