Distilled from production systems at BigTech — practical patterns, trade-offs, and measurement.
Fischer, Lynch, and Patterson's 1985 proof that consensus is impossible in async systems with even one crash fault — bivalent configurations, valency, and why waiting forever is the only option.
Leader elections, replicated state machines, and the safety rules that stop two leaders from committing different logs — with an interactive 5-node cluster.
N, W, R, and quorum intersection — how Dynamo-style systems trade consistency strength for availability, with an interactive calculator.
Lamport timestamps, causal ordering, and the exact moment concurrent updates become conflicts — with vector clock diagrams and code.
CAP tells you what breaks during a partition — PACELC tells you what you trade every write the other 99.9% of the time.
Crash-stop vs crash-recovery vs Byzantine — why surviving a crash costs 2f+1 and surviving a liar costs 3f+1.
Physical clocks drift and NTP steps — Lamport and hybrid clocks restore causal order without wall time.
The impossibility of perfect agreement over a lossy link — and why quorums exist. Interactive dropped-ACK demo.
Deterministic replicas must apply identical commands in identical order — one reordered message creates two databases.
A practical guide to 32-bit XID limits, MVCC tuple freezing, autovacuum parameters, production monitoring queries, and emergency recovery.