AP = AP in the CAP theorem, which states that in a distributed system you must either choose consistent and Partition tolerant (CP) or Available and Partition tolerant (AP), sort of the "heisenberg uncertainty theorem" of distributed system (you can't have CA because for practical purposes the "distributed" constraint implies "partition tolerant").
HA = "Always available". Answers may be stale, but you'll always get an answers.
CP = "Always consistent". You may not always get an answer, but if you do, it's consistent.
You can't have both in a distributed system, because distributed systems are on networks, networks are imperfect, and therefor you get "partitions' (a silly term for failures). The question is, given one of these failures, do you block and wait for all nodes to agree on what's "right", or do you just respond with the best information you have available.
CRDTs play an interesting role here. http://swarmdb.net/articles/acid2/
If I crash, I’m out. I can’t interfere with anything or anybody for long. If others think I crashed but I’m still running, I can become some sort of ghost haunting the system. Paranormal activity in an environment built entirely on logic is chaos.
A system built for partitions has to consider the problem where I’m talking but nobody can hear me, what do I do to sort it out, and what can I productively do until then, like queue up small purchases until my connection to Visa is available again.
CP : Consistent under partitioned conditions (when a n/w partition occurs in a distributed system you may not get answers for all your queries, but when you do get answers, they are not stale and are causally consistent)
These are concerns that are of particular interest when running a set of servers that communicate with each other directly, which is common in Elixir/Erlang due to this being supported by the OTP libraries.
The AP comes from CAP theorem, which is better documented on wikipedia than I could give justice to.