Abstract
The protocols underlying large-scale geo-replicated storage systems implement a wide spectrum of transactional consistency criteria. These protocols typically execute in two phases: ● In the Execution phase, a transaction reads from a database snap-shot and buffers its updates. ● In the Termination phase, a multicast/2PC protocol commits updates atomically while performing a so-called certification check that enforces the desired consistency guarantees. In this talk, we study various snapshotting algorithms for the Execution phase, and how they interact with the Termination phase. We propose a new approach, called Physics, that relies on a single scalar taken from a physical clock for tracking causal dependencies and building causally-consistent snapshots. We combine Physics with di erent certification algorithms, in order to implement several well-known consistency criteria, i.e., Causal Consistency, Non-Monotonic Snapshot Isolation, and Update Serialisability. We compare the resulting protocols to existing systems implementing similar guarantees.