The hitchhiker's guide to distributed transactions

Recent years have seen a flurry of research, both from academia and industry, enabling distributed transactions while minimizing the performance overhead. There are various proposals afoot; some systems constrain the transaction model, others will trade-off behavior under contention. Some systems sacrifice isolation guarantees, while others will incur higher read latencies.

This talk is a survey of the various approaches that academic (Carousel, MDCC, SLOG, TAPIR) and industrial (Spanner, CockroachDB, OceanVista) systems use to achieve atomicity in their transactions. We'll explore how each system uniquely composes its transaction model with the underlying replication protocol, each achieving the theoretical minimum latency for atomic commitment: one round-trip between data centers.