REST Architectural Style – Compliance of Archetype Controller

I think objects are only built correctly around coherent behaviors and not around data. I will provoke and say that data is almost irrelevant in the object oriented world. In fact, it's possible and sometime common to have objects that never return data, for example "log sinks", or objects that never return the data they are passed, for example if they calculate statistical properties.

Let us not confuse the PODS, (which are little more than structures), and the real objects that have behaviors (like the Contacts class in your example)¹.

PODS are basically a convenience used to talk to repositories and business objects. They allow the code to be type safe. No more, no less. Business objects, on the other hand, provide concrete behaviors, like validating your data, or storing it, or using it to perform a calculation.

So, behaviors are what we use to measure "cohesion"², and it's easy enough to see that in your object example there is some cohesion, even though you only show methods to manipulate top level contacts and no methods to manipulate addresses.

Regarding REST, you can see REST services as data repositories. The big difference with object oriented design is that there is (almost) only one design choice: you have four basic methods (more if you count HEAD, for example) and of course you have a lot of leeway with the URIs so you can do nifty stuff like pass many ids and get a larger structure back. Do not confuse the data they pass with the operations they perform. Cohesion and coupling are about code and not data.

Clearly, REST services have high cohesion (every way to interact with a resource is in the same place) and low coupling (every resource repository does not require knowledge of the others).

The basic fact remains, though, REST is essentially a single repository pattern for your data. This has consequences, because it's a paradigm built around easy accessibility over a slow medium, where there is a high cost for "chattiness": clients typically want to perform as few operations as possible, but at the same time only receive the data they need. This determines how deep an data tree you are going to send back.

In (correct) object oriented design, any non-trivial application will do much more complex operations, for example through composition. You could have methods for doing more specialized operations with the data -- which must be so, because while REST is an API protocol, OOD is used to build entire user facing applications! This is why measuring cohesion and coupling is fundamental in OOD, but almost insignificant in REST.

It should be obvious by now that analyzing data design with OO concepts is not a reliable way to measure it: it's like comparing apples and oranges!

In fact, it turns out that the benefits of being RESTful are (mostly) the ones outlined above: it's a good pattern for simple APIs over a slow medium. It is very cacheable, and shardable. It has fine grained control over chattiness, etc.

I hope this answers your (quite multifaceted) question :-)

^{1 This problem is part of a larger set of problems known as Object-Relational impedance mismatch. Proponents of ORMs are generally in the camp that explores the similarities between data analysis and behavior analysis, but ORMs have fallen under criticism of late because they seem not to really resolve the impedance mismatch and are considered leaky abstractions.}

^{2 http://en.wikipedia.org/wiki/Cohesion_(computer_science)}

You asked:

Does local applications running on the same machine that host the server, will have faster REST communication?

Absolutely. Your data won't have to be serialised onto the wire.

For faster "local" REST communications, does a server needs to expose 2 endpoints (one for localhost / one for remote) ?

No. Your server will expose on a particular address, and your client can either talk to localhost (127.0.0.1) or the actual server IP address. In the latter case your client can work on a remote host without reconfiguration, but fundamentally your server won't care whether your client is local or remote.

(there are perhaps subtleties in the above relating to multiple network interfaces, network stack optimisations etc., but they don't affect the underlying principles)

If you can colocate your two services, that may make sense. As soon as you have two machines and a bit of wire inbetween, then you're likely introducing complexity and a degree of unreliability. If you can colocate, and you accept that taking that one machine down will knock out both services, that's likely going to be a more manageable solution for you. Note that colocating can have adverse effects if (say) your machine doesn't have sufficient CPU/memory resources for running both processes.

Note that distributing across a network leaves you open to the Fallacies of Distributed Computing. Some of these fallacies will still apply in your case, but colocating will reduce your exposure.