Which Consistency Guarantee Do You Actually Need? — Linearizable, Weak, or Eventual Diagram

The Cost of Linearizability

Every read and write pays a quorum round-trip: a response is only as fast as the slowest replica that must answer
Attiya-Welch: the latency of a linearizable operation can never beat the uncertainty of the network delay between replicas
During a network partition, replicas that cannot reach a quorum must refuse requests — availability is traded away for correctness

When Weaker Guarantees Are Enough

Read-your-own-writes and monotonic reads cover the experience of one user; consistent prefix keeps causally related writes in order
Lamport timestamps and hybrid logical clocks give every event a version number that respects causal order
Eventual consistency — multi-leader or leaderless — lets clients write while offline and sync later

Cross-Channel Timing and Hard Constraints

When a message queue and a storage system both carry one fact, a reader can see them disagree unless one channel orders both
Reading through the same channel that accepted the write gives read-your-writes without global consensus
Uniqueness, non-negative balances, and the last remaining seat are hard constraints: guard them with consensus-backed compare-and-set, or a single-node timestamp oracle