I have been thinking about tiered storage wherein infrequently used data can be moved off to slow (cold) storage (like S3). I think that CEP-17 in conjunction with CEP-21 provides an opportunity for an interesting approach.
As I understand it CEP-17 clarified the SSTables interface(s) so that alternative implementations are possible, most notably CEM-25 (trie format sstables). CEP-21 provides a mechanism by which specific primary key blocks can be assigned to specific servers. It seems to me that we could implement an SSTable format that reads/writes S3 storage and then use CEP-21 to direct specific keys to servers that implement that storage. I use primary key because I don't think we can reasonably partition the records onto cold storage using any other method. I think that records on the cold storage may be deleted, and may be updated but both operations may take significant time and would require compaction to be run at some point. I expect that compaction would be very slow. I am certain there are issues with this approach and am looking for feedback before progressing an architecture proposal. Thanks, Claude