期刊名称：Bulletin of the Technical Committee on Data Engineering
出版年度：2016
卷号：39
期号：1
页码：27
出版社：IEEE Computer Society
摘要：Today’s replicated transactional storage systems typically have a layered architecture, combiningprotocols for transaction coordination, consistent replication, and concurrency control. These systemsgenerally require costly strongly-consistent replication protocols like Paxos, which assign a total orderto all operations. To avoid this cost, we ask whether all replicated operations in these systems need to bestrictly ordered. Recent research has yielded replication protocols that can avoid unnecessary ordering,e.g., by exploiting commutative operations, but it is not clear how to apply these to replicated transactionprocessing systems. We answer this question by analyzing existing transaction processing designs interms of which replicated operations require ordering and which simply require fault tolerance. Wedescribe how this analysis leads to our recent work on TAPIR, a transaction protocol that efficientlyprovides strict serializability by using a new replication protocol that provides fault tolerance but notordering for most operations.
