Agreement is bought with assumptions

A subject the papers are about. The loosest grouping, and the one to reach for last.

Whether a set of processes can agree on anything, and what has to be assumed before they can. Every positive result here is a bound with conditions attached, and the negative one says why the conditions cannot all be dropped.

Two records come first because agreement presupposes them. Lamport's happened-before relation gives a distributed system an ordering without a shared clock; Chandy and Lamport's snapshots give it a global state nobody is in a position to observe directly. Neither is about agreement, and neither can be skipped: you cannot agree on a value without first saying what it would mean for the system to be in a state at all.

Then the bounds. Pease, Shostak and Lamport establish what it costs to reach agreement when some participants are faulty, and the Byzantine Generals paper is the same result told as the story that gave the failure model its name. Fischer, Lynch and Paterson supply the impossibility: no deterministic protocol reaches consensus in an asynchronous system if even one process may fail. That result does not defeat the others -- it prices them, by showing exactly which assumption, synchrony, they were quietly buying.

The two later records argue about whether the price is right. Gaertner's *arbitrary is not (always) random* says the Byzantine model is often more pessimistic than the setting warrants, so protocols pay for adversaries they do not face. Rodrigues and Liskov take the opposite tack for systems that must run for years, where an adversary has time to corrupt replicas one by one and the model is not pessimistic enough.

FLP is held twice, the 1983 conference paper and the 1985 JACM version, deliberately. Several members sit in `who-paid-for-the-foundations`, which is not a coincidence: this is 1978-85 American defence-funded work, and the fault models are shaped accordingly.