Consensus Underground: The Failure Models Behind Raft, PBFT, and Proof-of-Stake

This book is for backend and infrastructure engineers who need to reason about consensus as a systems-design choice, not just recognize algorithm names. It starts from the practical questions the algorithms must answer: when silence can mean either a crash or delay; when a node may lie; what changes when participant identities cannot be counted; and where ordering, commitment, finality, data availability, and external side effects actually meet. Raft, PBFT, and Proof of Stake are not interchangeable voting schemes. Raft assumes known nodes that may stop but do not lie. PBFT thickens quorum requirements to survive Byzantine behavior. Permissionless PoW and PoS systems must also solve identity and Sybil resistance, and may expose different finality and economic assumptions. Following those differences helps readers judge which guarantees apply—and which do not—before choosing or operating a design. Across four chapters, the book connects Paxos, Raft, Viewstamped Replication, Zab, EPaxos, Multi-Raft, PBFT, Tendermint, HotStuff, DAG-BFT, PoW, PoS, Avalanche, CRDTs, Gossip, and SWIM. It also covers production concerns the papers leave outside the core algorithm, including persistence ordering, snapshots, membership changes, backpressure, linearizable reads, fencing tokens, and request deduplication. By the end, readers can compare designs by failure, network, membership, trust, finality, scale, and operational requirements—and recognize when a problem does not need consensus at all. The discussion uses concise diagrams and pseudocode and explains the quorum reasoning as it goes. Familiarity with basic backend concepts such as processes, networks, storage, and replicated state is helpful; no formal proof background is required.