Networking Technologies for Distributed Databases: From TCP/IP to RDMA and Programmable Fabrics

Abstract

Distributed databases are networked systems: every remote read, replicated write, quorum decision, failure-detection event, and reconfiguration crosses a communication substrate whose latency, throughput, loss behavior, and isolation properties shape database performance and availability. This survey organizes the supporting technologies into five layers---data-center and wide-area fabrics, transports, RPC and kernel-bypass mechanisms, acceleration and virtualization, and database coordination protocols. It compares TCP, QUIC, Ethernet, InfiniBand, RoCE, RDMA, user-space RPC, overlays, TLS, clock synchronization, and programmable switches, then relates them to representative systems including Spanner, CockroachDB, Aurora, FaRM, FaSST, and Raft-based stores. The central finding is that no network technology is uniformly best: reliable TCP-based stacks maximize portability and operational maturity; RDMA and kernel bypass minimize CPU and latency within controlled clusters; and WAN deployments benefit most from locality, topology-aware quorums, batching, and careful consistency choices. The survey concludes with selection guidance and research challenges involving tail latency, mixed lossless/lossy fabrics, multi-tenancy, observability, and safe in-network computation.