A Practical Analysis of the Gorums Framework: A Case Study on Replicated Services with Raft

Abstract

Gorums is a novel RPC framework developed to make it easier to build fault tolerant distributed systems. We want to assess whether Gorums can simplify the implementation of a practical fault tolerant service that supports reconfiguration. The Raft consensus algorithm is implemented in Gorums with the ability to do single-server configuration changes. In addition, we perform a background study of two state of the art Raft implementations. The abstractions used in these implementations are then compared to the abstractions Gorums provides and how they are used in our Raft implementation. A service is created that can be used with any of the aforementioned Raft implementations for consistency and fault tolerance. This service is then used to evaluate the different implementations through experimentation. Our evaluation shows that the Raft implementation that uses Gorums perform better with regards to latency and overall throughput during normal operation.

We do however discover this implementation to be sensitive to omission faults, which can further lead to availability issues if not handled properly. We solve this by developing extensions to Raft and Gorums. We show that these methods perform on a similar level when compared with the state of the art implementations. Results from our implementation efforts indicate that Raft's log replication process is problematic to implement with Gorums' abstractions. We discover that this is due to Raft adopting a monolithic design aimed to reduce the number of different RPC types, breaching the separation of concerns design principle.