CAP theorem

Cut the link between two servers and one must refuse or answer stale

The CAP theorem: when the network between replicas splits (a partition), a database can't be both Consistent and Available. Splits do happen, so the real choice is C or A while one lasts; a healthy network allows both.

Consistent means every read returns the latest write, as if there were only one copy. Available means every working server answers every request instead of returning an error.

Say the last concert ticket is stored in London and New York, and the cable between them is cut. Alice buys the ticket in London. New York can refuse Bob (consistent) or tell him "1 left" and maybe sell it twice (available).

CAP theorem

Amazon's shopping cart never refuses an "Add to cart"

Amazon's Dynamo paper (2007) describes a cart that accepts writes even when replicas can't reach each other, which is an AP choice. Conflicting versions are merged later, so an item you deleted can sometimes reappear.

Kubernetes chooses CP. It stores cluster state in etcd, which only accepts a write when a majority of its servers agree. A server that is cut off refuses writes rather than create a second version of the truth.

CAP theorem

Reads and writes must overlap, or each side answers on its own

Databases like Cassandra let you set how many of the N replicas must confirm a write (W) and how many must answer a read (R). If R + W > N, every read group shares a replica with the latest write, so reads see the newest value.

During a partition, that overlap costs availability: a side that can't reach enough replicas must refuse. With small R and W every side still answers, but reads can be stale. No setting escapes the trade.