Linearizability
Its the idea of appearing like there's only a single node while still having multiple replicas. And all operations looking like they are atomic. Whenever a write is done, all clients reading the service, as example a database, should be able to see the recent change immediately.
Linearizability and Serializability differ, because serializability is about sequential execution of concurrent transactions. Linearizability is just the idea of having a recency-guarantee when reading data;
Linearizability is kinda important for locking and leader election, e.g, when all nodes startup we need to ensure they are electing a single leader and not causing split-brain, or a unique constraint on a database.
How to implement it on your system?
What kind of Distributions Models are linearizable?
| Strategy | Works? |
|---|---|
| Single Leader Replication | Potentially |
| Consensus Algorithm | Yes! |
| Multi-Leader | Nope |
| Leaderless | Almost certainly not |
For leaderless, as example DynamoDB, theres a claim that with a quorum-reads and writes(w + r > N), we can reach a linearizable system, which is not exactly true - Only if you sacrifice performance requiring reads and writes to read the last state. Last-Write-Wins from Cassandra, fall into Clock Issues and Clock Skew and are almost certainly non linearizable.
Single Leader is only linearizable on a single datacenter deployment. Might not be solution for all, since with multi datacenters, it loses linearizability on networks failures. It is a know problem and consequence of Single or Multi-Leader with multiple datacenters deployment.