Alternative History: Multi-Log Systems in Theory and Practice [electronic resource].

9781085777049

Ann Arbor : ProQuest Dissertations & Theses, 2019.

1 online resource (94 p.)

Access is available to the Yale community.

Access restricted by licensing agreement.

Large-scale datacenter systems rely on stateful control plane services. Distributing state for such services can be difficult; idiosyncratic interfaces and performance requirements precludes the use of general, fixed API storage systems, while custom solutions are often complicated and difficult to debug and maintain. By funneling all updates through a single, durable, totally ordered log, Shared Log State Machine Replication can make arbitrary state available, durable, and strongly consistent among a number of machines using a single service. The log abstraction hides the complexity of asynchrony and failures from applications, allowing them to be built as simple state machines that process updates in strict sequence. However, this simplicity comes at a cost to speed and flexibility by imposing a system-wide total order that is expensive, often impossible, and typically unnecessary. Furthermore, the State Machine Replication paradigm itself functions as a broadcast domain, forcing servers to see more state than is actually useful. This work investigates the implications of weakening the Shared Log State Machine Replication semantics. By guaranteeing a partial ordering of updates instead of a total order, the FuzzyLog achieves linear scaling with transaction support, weak consistency models, and progress during network partitions. By allowing servers to filter which updates they see within the context of an object, FuzzyViews enable greater privacy and read performance.

Dissertations & Theses @ Yale University.

Books / Online / Dissertations & Theses

English

January 17, 2020

Thesis (Ph.D.)--Yale University, 2019.

Yale University. Computer Science.