Revising 1-Copy Equivalence in Replicated Databases with Snapshot Isolation

Multiple database replication protocols have used replicas supporting the snapshot isolation level. They have provided some kind of one-copy equivalence, but such concept was initially conceived for serializable databases. In the snapshot isolation case, due to its reliance on multi-versioned concurrency control that never blocks read accesses, such one-copy equivalence admits two different variants. The first one consists in relying on sequential replica consistency, but it does not guarantee that the snapshot used by each transaction holds the updates of the last committed transactions in the whole replicated system, but only those of the last locally committed transaction. Thus, a single user might see inconsistent results when two of her transactions have been served by different delegate replicas: the updates of the first one might not be in the snapshot of the second. The second variant avoids such problem, but demands atomic replica consistency, blocking the start (i.e., in many cases, read accesses) of new transactions. Several protocols of each kind exist nowadays, and most of them have given different names to their intended correctness criterion. We survey such previous works and propose uniform names to these criteria, justifying some of their properties.