Achieving both low latency and strong consistency at large scale
Speaker: Seo Jin Park
Since the advent of the Internet, applications have started using millions of people's data which is beyond the capacity of a single machine. In response to the need for large-scale systems, the systems community focused on largely-scalable distributed systems, such as NoSQL storage. Unfortunately, the proposed solutions for scalability introduced two new challenges for applications: higher latency to access data (which limits the form of computation) and weak consistency (which makes systems harder to use).
This talk will focus on how to improve the latency and consistency of large-scale systems, which are necessary to enable a new class of complex applications that operate on big data in real-time. First, I will discuss a consistency problem in Remote Procedure Calls (RPC). In the presence of failures, a single RPC invocation may result in multiple executions. I will present Reusable Infrastructure for Linearizability (RIFL) which prevents the consistency problem by guaranteeing exactly-once execution of RPCs [SOSP'15]. Secondly, replicated systems often compromise consistency for performance; consistent replication doubles latency and reduces throughput. I will present Consistent Unordered Replication Protocol (CURP), which removes the overheads of strongly-consistent replication [NSDI'19].
Bio: Seo Jin Park is a PhD candidate in computer science at Stanford University, advised by John Ousterhout. His research focuses on improving latency and consistency of large-scale distributed systems. He previously worked on a fast in-memory storage system (RAMCloud), and on a nanosecond-scale fast logging system (NanoLog). His PhD study was supported by a Samsung Scholarship. Before coming to Stanford for PhD, he previously received M.Eng. in EECS and B.S. in Computer Science and Mathematics from MIT in 2013.