The popular publish/subscribe communication paradigm, for building large-scale distributed event notification systems, has attracted attention from both academia and industry due to its performance and scalability characteristics. While ordinary "web surfers" typically are not aware of minor packet loss, industrial applications often have tight timing constraints and require rigorous fault tolerance. Some past research has addressed the need to tolerate node crashes and link failures, often relying on distributing the brokers on an overlay network. However, these solutions impose significant complexity both in terms of implementation and deployment.

In this paper, we present a crash tolerant Paxos-based pub/sub (P2S) middleware. P2S contributes a practical solution by replicating the broker in a replicated architecture based on Goxos, a Paxosbased fault tolerance library. Goxos can switch between various Paxos variants according to different fault tolerance requirements. P2S directly adapts existing fault tolerance techniques to pub/sub, with the aim of reducing the burden of proving the correctness of the implementation. Furthermore, P2S is a development framework that provides sophisticated generic programming interfaces for building various types of pub/sub applications. The flexibility and versatility of the P2S framework ensures that pub/sub systems with widely varying dependability needs can be developed quickly.

We evaluate the performance of our implementation using event logs obtained from a real deployment at an IPTV cable provider. Our evaluation results show that P2S reduces throughput by as little as 1.25% and adds only 0.58 ms latency overhead, compared to its non-replicated counterpart. The performance characteristics of P2S prove the feasibility and utility of our framework.

Building robust network services that can withstand a wide range of failure types is a fundamental problem in distributed systems. The most general approach, called Byzantine fault tolerance, can mask arbitrary failures. Yet it is often considered too costly to deploy in practice, and many solutions are not resilient to performance attacks. To address this concern we leverage two key technologies already widely deployed in cloud computing infrastructures: replicated state machines and intrusion

detection systems.

First, we have designed a general framework for constructing Byzantine failure detectors based on an intrusion detection system. Based on such a failure detector, we have designed and built a practical Byzantine fault-tolerant protocol, which has costs comparable to crash-resilient protocols like Paxos. More importantly, our protocol is particularly robust against several key attacks such as flooding attacks, timing attacks, and fairness attacks, that are typically not handled well by Byzantine fault masking procedures.