Practical Byzantine Fault Tolerance
Castro and Liskov present a practical state-machine-replication protocol tolerating f Byzantine replicas among 3f+1, with normal-case ordering, checkpoints, view changes, and recovery.
PBFT's normal path assigns requests with PRE-PREPARE, establishes a prepared certificate with PREPARE, and makes that knowledge durable across view changes with COMMIT. More than matching commits permit execution in a system of replicas. Checkpoints, watermarks, client timestamps, state transfer, and view change turn the core quorum argument into an operable service.
The reported NFS benchmark is evidence that Byzantine replication was practical on its 1999 hardware and network. It is not a present throughput claim, and later protocols alter authentication, leaders, pipelining, and communication patterns.
Visual audit #
All 14 physical pages were rendered and inspected. The architecture, normal-case message flow, and performance plots are information-bearing. GTK uses an independently drawn sequence focused on the protocol invariant; the historical benchmark plots are not retained because their hardware-specific magnitudes do not guide current capacity.