Comment by saurik

I don't understand why every distributed project has gravitated to write-optimized data structures; like, if you look at all of the decentralized SQL offerings, they are now all built around LSN trees... does no one have extremely large transactional (so not like, bulk data analysis OLAP but standard OLTP work) datasets that are more read than write (ones where raw replication to balance the I/O or CPU load of queries due to the space exploitation, so you want the kind of autosharding these solutions offer)? I guess I am just confused as I would have expected write-oriented workloads to be more rare and specialized, mostly operating in the regime of transient shopping carts (Amazon's original use case for Dynamo) or click logs or whatever, as I would have thought most projects (maybe a social network or a chat client) would have numerous reads (maybe even tens of thousands) for every write.