Consistency Models

Consistency Models

Step 1 of 10

copies3

One key, three copies. The interesting question is not where it is stored — it is what a read is allowed to return.

The Idea

Replication hands every node a copy, so the real question is what a read may return. Strong consistency always shows the newest committed write. Eventual consistency may show any recent copy. Between them sit the guarantees users actually notice: read-your-writes, and causal order.

Real-World Example

You post a comment on a social feed, refresh, and it is gone. Your write landed on one replica and your read hit another. Read-your-writes pins the session to a replica that has your write, so the page stops contradicting you.

The Tradeoff

Strong reads cost a coordination round trip and stall during a failover. Weaker models are cheap and stay up, but hand the anomalies to your application. Choose the weakest model whose anomalies a user would never notice, then write that choice down where the next team will find it.

Your turn

Put the steps in the right order.

1. 1The client is routed to a replica that has not applied it yet↑↓
2. 2A write is accepted by the leader and acknowledged↑↓
3. 3Later reads carry that version, so a lagging replica waits or forwards↑↓
4. 4The refresh comes back missing the comment the user just posted↑↓

Check order

Mini quiz

1 / 3

A read-your-writes guarantee promises that:

• AEvery client sees writes in the same order
• BYou never read a version older than your own last write
• CAll replicas hold identical bytes at all times
• DWrites are applied in timestamp order