Partition tolerance in CAP means tolerance to a network partition. An example of a network partition is when two nodes can't talk to each other, but there are clients able to talk to either one or both of those nodes. If you've ever used IRC and experienced a netsplit, this is a great example of that.

A CA system guarantees strong consistency, at the cost of not being able to process requests unless all nodes are able to talk to each other. An AP system is able to function during the network split, while being able to provide various forms of eventual consistency.

There's several forms of eventual consistency. There's the "weak eventual consistency", where you may not be able to read your writes. In the context of Dynamo-based key/value stores this actually only occurs as a failure condition: when the first node in the preference list for a key isn't available.

Another form of eventual consistency is "read-your-writes" consistency, where by you use a quorum of R reads, W writes out of a total of N replicas and set R + W > N. In this case you're guaranteed to be able to read your writes even if multiple nodes in the preference list for a key fail (as long as you're able to meet a quorum).