CAP Theorem

In the year 2000, Eric Brewer stood up at the ACM Symposium on Principles of Distributed Computing and made a conjecture that would reshape how every major tech company designs systems. Two years later, Seth Gilbert and Nancy Lynch proved it mathematically.

The conjecture was deceptively simple: a distributed data store cannot simultaneously provide more than two of three guarantees. Consistency, Availability, and Partition Tolerance.

Now, every blog post and interview prep site repeats that sentence. But most of them get the implications wrong. They treat it like a menu where you pick two items. "I'll take consistency and availability, hold the partition tolerance." That's not how it works. Not even close.

Here's the uncomfortable truth: network partitions will happen. Your data center will lose connectivity to another data center. A switch will fail. A cable will get cut. Partitions are not a theoretical concern, they are an operational certainty. So partition tolerance isn't optional. You must pick P. The actual choice is between C and A during a partition.

This lesson goes deep. Not the interview-prep version where you memorize "CP vs AP." The real version, where you understand why Google built Spanner, why Amazon built DynamoDB, and why the CAP theorem is simultaneously the most important and most misunderstood result in distributed systems.

Most explanations of CAP are hand-wavy. Let's be precise about what each property actually means, because the devil is entirely in the definitions.

Consistency (C). Every read receives the most recent write or an error. This is linearizability, the strongest form of consistency. If client A writes value "42" and the write completes successfully, then any subsequent read by any client, on any node, must return "42" or something newer. Not "42 eventually." Not "42 on most nodes." Every read, every node, immediately.

Availability (A). Every request received by a non-failing node must result in a response. Not "a response within 10 seconds." Not "a response from the primary." Every non-failing node must respond. If a node is up and receives a request, it cannot just sit there waiting for another node to come back online. It must respond.

Partition Tolerance (P). The system continues to operate despite an arbitrary number of messages being dropped or delayed between nodes. A network partition splits your cluster into two groups that can't communicate. The system must keep functioning in some capacity.

Here's the critical insight most people miss: you don't choose between C, A, and P as a permanent architectural decision. You make this choice per operation and during partitions. When the network is healthy, you can have all three. The theorem only constrains you when a partition actually occurs.