CAP Theorem: Pick Two of Three Guarantees

Why it exists

Distributed systems must function over unreliable networks where messages can be lost or delayed. The CAP theorem was formulated to explain the fundamental trade-offs a system must make when its nodes can't communicate. It provides a clear vocabulary for reasoning about system design choices under network failure conditions.

The mental model

Imagine a shared bank account ledger accessible by two tellers in different cities. If the phone line between them goes down (a network partition), they have a choice. To guarantee consistency, they could stop all transactions until the line is restored, ensuring the balance is never wrong. This is a CP system (Consistency over Availability). Or, to stay available, they could allow transactions based on the last known balance, risking an overdraft if both process a withdrawal. This is an AP system (Availability over Consistency). They cannot do both simultaneously.

How it works

The theorem defines three guarantees. Consistency means every read receives the most recent write or an error. Availability means every request gets a non-error response, though it may not be the most recent data. Partition Tolerance means the system continues to operate despite network failures between nodes. The theorem proves that in the event of a partition, a system must choose to sacrifice either Consistency or Availability. You can't have all three.

When to use it

Use CAP as a framework when choosing or designing a distributed database. For systems where data integrity is non-negotiable, like financial transactions or inventory management, a CP system is appropriate. It will refuse requests during a partition to prevent inconsistency. This is the choice for systems that prioritize correctness above all else.

When not to use it

For systems where high uptime is more critical than perfect data consistency, an AP system is a better fit. Examples include social media feeds or analytics dashboards, where serving slightly stale data is an acceptable price for keeping the service online and responsive. A CA (Consistent, Available) system is largely theoretical for distributed systems, as it assumes a perfect network that never fails, which is unrealistic.

One canonical example

During a network partition, a CP database like Google's Spanner might refuse a write request or make a client wait if it cannot guarantee all replicas are updated. In contrast, an AP database like Amazon's DynamoDB will accept the write on an available node and resolve potential conflicts later, ensuring the user's request succeeds even if data is temporarily inconsistent across the cluster.