Ceph

Sam pointed out that we could turn our reads into essentially an Available (versus Consistent) model just by turning on read-from-replica, and that we can further handle a down node by requesting from a separate replica after a (very short) timeout.

The use cases for these kinds of features are very narrow, but RGW fits into the niche. It would enhance viability as a web-serving service, since we don't need to wait for full timeout detection and peering if an OSD dies in order to keep serving up object reads.