DDIA Chapter 9: Consistency and Consensus

Eventual consistency is easy to operate and easy to misuse. Stronger models cost latency or availability. Consensus is how nodes agree on leaders, order, and commits when intuition is not enough.

Linearizability

The system behaves as if there is one copy of the data and every operation is atomic (a recency guarantee). After any read sees x=2, all later reads must see that (or newer). Not the same as serializability (transaction isolation).

Single-leader + consensus can provide linearizability; multi-leader generally cannot; Dynamo-style only with painful caveats. CAP (informally): insist on linearizability during a partition and you sacrifice availability. Few systems are linearizable end-to-end because it is slow even on a healthy network.

Ordering and causality

Causal consistency respects happens-before and is the strongest model that need not stall on network delay the way linearizability does. Tools: version vectors, Lamport timestamps, total-order broadcast. Lamport gives compact total order but cannot always detect concurrency the way version vectors can.

Consensus and distributed transactions

Consensus: agree on a value (leader election, atomic commit). Impossible in fully async unreliable models in theory; practical with timeouts (Raft, Paxos, Zab, Viewstamped Replication).

Two-phase commit: prepare then commit; after a yes vote, participants may block if the coordinator dies. Operationally heavy; can freeze locks. Total-order broadcast ≈ repeated consensus and underpins many replication stories.

flowchart TD
  Eventual[Eventual] --> Causal[Causal]
  Causal --> Linear[Linearizable]
  Linear --> Consensus[Consensus / Raft]

Example

Snippets below use Java, SQL, or plain text as labeled.

Linearizability vs serializability:

Linearizability: after any read sees x=2, later reads see x>=2 (recency)
Serializability: transactions act as if run one-at-a-time (isolation)

Lamport bump:

long onReceive(long remoteTs, long localTs) {
  return Math.max(remoteTs, localTs) + 1;
}

Bottom line

Linearizability is about recency; serializability is about transactions. Causal consistency is often enough for geo systems. Consensus is powerful and sensitive to membership and timeouts.