Byzantine Fault Tolerance from the Fundamentals

Consensus in the Presence of Partial Synchrony

For me the most interesting part is model definition and the implementation of rounds. These ideas are used in all algorithms to create some sort of round structure, and is essential for termination. Consensus algorithms are not considered as most advanced but there are some nice ideas, like locking and unlocking concepts.

Practical byzantine fault tolerance and proactive recovery

Fast Byzantine Consensus

Regarding these two papers, although they are definitely seminal, they are not necessarily easy read, and they don't in a clear way express core mechanisms. Regarding PBFT, the most important idea (in my view) is implementing Byzantine-tolerant consensus without signatures. But this mechanism is hidden in the SMR algorithm. This might be useful to know in the context of quantum-resistant algorithms as only assumes secure point-to-point computation. Proactive recovery part has some cool ideas in the context of snapshotting and catch-up of algorithms. The other paper is simpler read, but I am not sure it also explains the idea as simple as possible, as for example it relies on the signatures which is not needed in case of 5f+1. Signature vs signature-free was big deal in the past as algorithms were mostly designed for LAN, and there the cost of signatures matters a lot."

Unifying Byzantine Consensus Algorithms with Weak Interactive Consistency

This is a paper I was involved that extract core consensus mechanism of both of these algorithms and present it in a simpler way using DLS round based model.

Automatic Reconfiguration for Large-Scale Reliable Storage Systems

In the context of automatic reconfiguration in case of BFT protocols, this is nice paper.