Geo-replicated data stores are one of the integral parts of today's Internet services. Service providers usually replicate their data on different data centers worldwide to achieve higher performance and data durability. However, when we use this approach, the consistency between replicas becomes a concern. At the highest level of consistency, we want strong consistency that provides the illusion of having only a single copy of the data. However, strong consistency comes with high performance and availability costs. In this work, we focus on weaker consistency models that allow us to provide high performance and availability while preventing certain inconsistencies. Session guarantees (aka. client-centric consistency models) are one of such weaker consistency models that prevent some of the inconsistencies from occurring in a client session. We provide modified versions of session guarantees that, unlike traditional session guarantees, do not cause the problem of slowdown cascade for partitioned systems. We present a protocol to provide session guarantees for eBay NuKV that is a key-value store designed for eBay's internal services with high performance and availability requirements. We utilize Hybrid Logical Clocks (HLCs) to provide wait-free write operations while providing session guarantees. Our experiments, done on eBay cloud platform, show our protocol does not cause significant overhead compared with eventual consistency. In addition to session guarantees, a large portion of this dissertation is dedicated to causal consistency. Causal consistency is especially interesting as it is has been proved to be the strongest consistency model that allows the system to be available even during network partitions. We provide CausalSpartanX protocol that, using HLCs, improves current time-based protocols by eliminating the effect of clock anomalies such as clock skew between servers. CausalSpartanX also supports non-blocking causally consistent read-only transactions that allow applications to read a set of values that are causally consistent with each other. Read-only transactions provide a powerful abstraction that is impossible to be replaced by a set of basic read operations. CausalSpartanX, like other causal consistency protocols, assumes sticky clients (i.e. clients that never change the replica that they access). We prove if one wants immediate visibility for local updates in a data center, clients have to be sticky. Based on the structure of CausalSpartanX, we provide our Adaptive Causal Consistency Framework (ACCF) that is a configurable framework that generalizes current consistency protocols. ACCF provides a basis for designing adaptive protocols that can constantly monitor the system and clients' usage pattern and change themselves to provide better performance and availability. Finally, we present our Distributed Key-Value Framework (DKVF), a framework for rapid prototyping and benchmarking consistency protocols. DKVF lets protocol designers only focus on their high-level protocols, delegating all lower level communication and storage tasks to the framework.
Read
