Hacker News Reader

A lot of things are stated as conclusions in this article, where SOTA has reversed or in some cases invalidated the conclusions. Unfortunately they are not published, and will probably remain trade secrets for another decade.

The biggest conclusion that is invalidated is that your archival workload cannot be bin packed with your hot workloads. With the ever reducing IO/byte of HDD, this has radically changed where the bytes go.

View on HN

Another example is the cost of IO when your backups are not perfectly coalesced. If you're striping writes across datacenters, into objects which are sharded across clusters that encode across many drives - your backups get extremely fragmented.

The IO for your backup suddenly becomes HUGE when compared to the size of the end object you wish to read. This makes things like tape nasty - sure you can read at incredible linear speeds.. but that's only worth it if you actually wanted to restore the exact TBs that are on the exact cartridge your drive & robot picked up.

Reply on news.ycombinator.com