A consistency model specifies what clients can expect to observe when they read and write data held on multiple replicas. It matters because a write accepted by one node does not automatically appear to every client at the same time. The central questions are order—which operations users see first—and visibility—when a change becomes visible. section 6.7
Two people edit one document
Alice adds a paragraph to a shared document. Shortly afterward, Bob changes another paragraph. Will everyone see Alice’s addition before Bob’s change? Will Bob see Alice’s addition immediately, or only after a delay? A consistency model tells clients which observations the system permits; it does not describe the edits themselves. section 6.7
Consistency guarantees vary in strength:
- Linearizability is the strong end described here: users see updates in the same order, with changes immediately visible to them despite replication. section 6.7
- Weaker models can permit delays in visibility or temporary differences in observed order. Eventual consistency is an example. section 6.7
Stronger guarantees carry a cost: coordinating copies of data adds overhead and can slow performance. Network partitions sharpen the trade-off between consistency and availability; even without a partition, consistency can compete with latency. Choose guarantees by asking what the application’s users may safely observe, rather than assuming every database makes the same promise. section 6.7 section 6.15
Keep three related ideas separate. Consensus concerns what distributed nodes agree on and can support strong consistency; a consistency model concerns what clients may observe. Isolation levels concern interactions between concurrent transactions. And ACID consistency concerns data integrity rules, not the visibility of updates across replicas. section 6.6 section 6.1.1






