The case for semantic aware remote replication

This paper argues that the network latency due to synchronous replication is no longer tolerable in scenarios where businesses are required by regulation to separate their secondary sites from the primary by hundreds of miles. We propose a semantic-aware remote replication system to meet the contrasting needs of both system efficiency and safe remote replication with tight recovery-point and recovery-time objectives. Using experiments conducted on a commercial replication system and on a Linux file system we show that (i) unlike synchronous replication, asynchronous replication is relatively insensitive to network latency, and (ii) applications such as databases already intelligently deal with the weak persistency semantics offered by modern file systems. Our proposed system attempts to use asynchronous replication whenever possible and uses application/file-system "signals" to maintain synchrony between the primary and secondary sites. We present a high-level design of our system and discuss several potential challenges that need to be addressed in such a system.