Computer infrastructures are inevitably subject to failures. The systems they host must therefore be designed to continue functioning despite these faults. Replication is a fundamental mechanism to achieve this goal: by maintaining multiple copies of data across distinct physical components, a distributed system can tolerate the failure of some of them. Designing correct replication mechanisms, however, remains challenging, as it requires carefully coordinating unreliable components and, in some cases, protecting against malicious behavior.

Modern data centers further complicate this task by challenging the assumptions on which distributed protocols are traditionally built. Advances in networking technologies, such as remote direct memory access (RDMA), together with the emergence of disaggregated architectures, enable sub-microsecond communication and resource pooling, giving rise to highly efficient, extremely low-latency systems. At the same time, one-sided memory access makes traditional message-passing protocols suboptimal, while disaggregation introduces new failure modes in which memory and compute can fail independently. These developments call for a re-examination of fundamental distributed computing problems in light of very low latency, remote memory access, and disaggregated resources.

This thesis addresses these challenges by designing, implementing, and evaluating novel fault-tolerance primitives tailored to the properties of modern data centers. Our work explores a two-dimensional design space spanning both replication mechanisms and fault models. First, we consider both read-write replication, which is particularly suited to key-value stores, and state-machine replication based on consensus, which provides a universal replication primitive. Second, beyond crash faults, we study arbitrary faults, known as Byzantine failures, reflecting the growing threat posed by malicious actors in multi-tenant infrastructures.

Our contributions are threefold. First, we introduce SWARM, a wait-free read-write replication protocol for disaggregated memory that provides strongly consistent access in a single memory round trip. Second, we present uBFT, a Byzantine fault-tolerant consensus engine that executes requests in under 10 microseconds and leverages the separation between compute and memory to operate with a small number of replicas. Third, we propose DSig, a hybrid signature scheme that enables signing, transmitting, and verifying a message in under 8 microseconds.

Together, our results demonstrate that strong fault tolerance need not come at the cost of high latency. By exploiting the properties of modern data center architectures, this thesis revisits classical fault-tolerance problems and shows how to build systems that are both reliable and extremely fast.