On limitations of using cloud storage for data replication

Cloud storage services often provide a key-value store (KVS) functionality, an object-based interface for accessing a collection of unstructured data items or blobs. Every blob is associated with a key that serves as identifier to access the blob. In the simplest form, a key-value store provides only methods for writing and reading an entire blob, for removing blobs, and for listing all defined keys. On the other hand, many existing schemes for replicating data with the goal of enhancing resilience (e.g., based on quorum systems) associate logical timestamps with the stored values, in order to distinguish multiple versions of the same data item. This paper uses the consensus number of a shared storage abstraction as a measure for its power to facilitate the implementation of data replication. It is demonstrated that a KVS is a very simple primitive, not different from read/write registers in this sense, and that a replica capable of the typical operations on timestamped data is fundamentally more powerful than a KVS. Hence, data replication schemes over storage providers with a KVS interface are inherently more difficult to realize than replication schemes over providers with richer interfaces.