Abstractions for the Stateful Control Plane

Seminar | October 26 | 12:30-1:30 p.m. | Soda Hall, 430 (Woz)

 Mahesh Balakrishnan, Facebook/Yale University

 Electrical Engineering and Computer Sciences (EECS)

At the heart of many cloud-scale systems is a logically centralized control plane that requires strong consistency and fault tolerance: examples include coordination services, SDN controllers, filesystem namespaces, and big data schedulers. Today, these control plane services are difficult to build, harden and scale, requiring complex protocols like Paxos and 2-phase commit that are inefficient when layered and difficult to combine. The shared log approach simplifies such applications by providing a data-centric abstraction that hides the complexity of the underlying distributed system. First generation shared log systems achieve this simplicity by imposing a global total order over updates; such systems have seen significant deployment in industry. In this talk, I'll describe the genesis, evolution, and current state of shared log systems, including new work that significantly improves their scalability and robustness.