Linearizability

Linearizability makes operations across several replicas appear to take effect once in a single real-time order. After a write finishes, any later read must not return the older value.

Linearizability and serializability differ because serializability concerns the equivalent sequential execution of concurrent transactions. Linearizability is just the idea of having a recency-guarantee when reading data;

Linearizability matters for locking and leader election. It prevents split brain, where two machines both believe they are the leader, and helps enforce a unique database value.

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
In leaderless systems, a quorum means contacting enough replicas for reads (r) and writes (w) that their sets overlap (w + r > N, where N is the replica count). Overlap alone does not guarantee linearizability: a read must still observe the latest completed write under failures and delays.
Last-write-wins conflict resolution depends on timestamps and therefore inherits clock-skew risks; it is generally not linearizable.

A single leader can offer linearizable behavior only while clients reliably reach the current leader and its replication rules preserve that guarantee. Multi-datacenter failures make that coordination harder; they do not automatically make a design non-linearizable.