Evgenii Ivanov's deep dive on The Consensus walks through quorum replication with runnable Python code, starting from Thomas's 1979 majority-voting scheme for replicated databases and Gifford's same-year extension that also routed reads through quorums. Because any two majorities intersect, conflicting updates cannot both be accepted, and with n replicas a system tolerates floor(n/2) unavailable ones — which is why replication factors are typically odd: three and four replicas both only tolerate one failure.

The article builds a simple quorum register in Python (Register, Replica, TGCluster classes) where writes carry a timestamp of logical clock plus replica ID, and reads return the highest-timestamped value from a majority. It then demonstrates the classic flaw, familiar from Designing Data-Intensive Applications: a delayed write can let one reader see the new value while a later reader still sees the old one. There is no commit notion — a replica exposes a value the moment it arrives, loosely analogous to Read Uncommitted — so the simple algorithm is not linearizable.

The ABD algorithm (Attiya, Bar-Noy, Dolev, 1990s) fixes this with a second read phase: after finding the newest value, the reader writes it back to a quorum before returning, so later reads cannot go backwards. Lynch & Shvartsman's 1996 MWABD extension adds multi-writer support via a timestamp-query phase before writing. The cost is an extra network round trip, which the article contrasts with leader-based protocols like Raft or Multi-Paxos that replicate a command in one round trip once a leader exists.

Two hard limits follow. First, ABD is not compare-and-swap: replicas are monotonic in timestamps, so a stale write can still collect a majority of ACKs while being silently ignored — demonstrated in code where a late write 'succeeds' and vanishes. CAS depends on a global ordering decision about the current value, which ABD never makes; a CAS1/CAS2 scenario shows CAS2 succeeding by consuming a value CAS1 disowned. Second, ABD is not consensus: it collapses the past into the latest timestamped value, so after crashes it can recover the newest register value but cannot distinguish committed log entries from partial writes. Murat Demirbas's verdict is quoted: ABD is 'memoryless and hedonistic.'

The practical echo is Cassandra: its blocking read repair mirrors ABD's write-back for monotonic quorum reads, while CAS-like operations require lightweight transactions built on Paxos — the same boundary in production form. The piece closes by recommending the book 'Quorum Systems With Applications to Storage and Consensus' for a formal treatment.