Services communicate to each other through an HAproxy instance on each host that is itself managed and configured by Sidecar. It is inspired by Airbnb's SmartStack. But, we believe it has a few advantages over SmartStack:

Native support for Docker (works without Docker, too!);
No dependence on Zookeeper or other centralized services;
Peer-to-peer, so it works on your laptop or on a large cluster;
Static binary means it's easy to deploy, and there is no interpreter needed;
Tiny memory usage (under 20MB) and few execution threads means its very light weight

With an AP system, you are giving up consistency, and not really gaining anything in terms of effective availability, the type of availability you really care about.  Some might think you can regain strong consistency in an AP system by using strict quorums (where the number of nodes written + number of nodes read > number of replicas).  Cassandra calls this “tunable consistency”.  However, Kleppmann has shown that even with strict quorums, inconsistencies can result.10  So when choosing (algorithmic) availability over consistency, you are giving up consistency for not much in return, as well as gaining complexity in your clients when they have to deal with inconsistencies.

In his excellent blog post [...] Jeff Hodges recommends that you use the CAP theorem to critique systems. A lot of people have taken that advice to heart, describing their systems as “CP” (consistent but not available under network partitions), “AP” (available but not consistent under network partitions), or sometimes “CA” (meaning “I still haven’t read Coda’s post from almost 5 years ago”).

I agree with all of Jeff’s other points, but with regard to the CAP theorem, I must disagree. The CAP theorem is too simplistic and too widely misunderstood to be of much use for characterizing systems. Therefore I ask that we retire all references to the CAP theorem, stop talking about the CAP theorem, and put the poor thing to rest. Instead, we should use more precise terminology to reason about our trade-offs.

I went into one of the instances and quickly did an iptables DROP on all packets coming from the other two instances.  This would simulate an availability zone continuing to function, but that zone losing network connectivity to the other availability zones.  What I saw was that the two other instances noticed the first server “going away”, but they continued to function as they still saw a majority (66%).  More interestingly the first instance noticed the other two servers “going away”, dropping the ensemble availability to 33%.  This caused the first server to stop serving requests to clients (not only writes, but also reads).

ZooKeeper, while tolerant against single node failures, doesn't react well to long partitioning events. For us, it's vastly more important that we maintain an available registry than a necessarily consistent registry. If us-east-1d sees 23 nodes, and us-east-1c sees 22 nodes for a little bit, that's OK with us.

Similar to ACID properties, if you partially provide properties it means the user has to _still_ consider in their application that the property doesn't exist, because sometimes it doesn't. In you're fsync example, if fsync is relaxed and there are no replicas, you cannot consider the database durable, just like you can't consider Redis a CP system. It can't be counted on for guarantees to be delivered. This is why I say these systems are hard for users to reason about. Systems that partially offer guarantees require in-depth knowledge of the nuances to properly use the tool. Systems that explicitly make the trade-offs in the designs are easier to reason about because it is more obvious and _predictable_.