Distributed Multi-Level Recovery in Main-Memory Databases

The authors present two schemes for concurrency control and recovery in distributed main-memory databases. In the client-server scheme, clients ship log records to the server, which applies the updates to its database copy. In the shared disk scheme, each site broadcasts its updates to other sites. The above enable the schemes to support concurrent updates to the same page at different sites. Both schemes support an explicit multi-level recovery abstraction for high concurrency, reduced disk I/O by writing only redo log records to disk during normal processing, and use of per-transaction redo and undo logs to reduce contention. Further, they use a fuzzy checkpointing scheme that writes only dirty pages to disk, yet minimally interferes with normal processing, not requiring updaters to even acquire a latch before updating a page.