I have recently had discussions with some colleagues about what architecture they prefer and, while they seem to favor thinly sliced services, I have come to the conclusion that the overhead that comes with slicing services thin is not worth the extra time that it takes to setup, verify and test the complex, internal communication that comes with this kind of architecture. Fat is better!
If I am designing a system that should work in a coherent way, I want it all in my big, fat, juice object model. This enables me to put the functionality where it is most cohesive and, therefore, gives me the best design possible. Every object should carry its own weight.
If there are external services they must, by necessity, be outside the model, but the internal representation of the external service should be inside my model.
An example of an architecture that relies on thinly sliced services is REST. REST is very elegant and it definitely has a place when publishing resources. But REST models are anemic. They rely on you to GET the information from the resource, do things to it and then replace the information of the resource with a PUT. It is CRUD for the web. It is not intended to take advantage of what is good in object-oriented and functional programming, like sending behaviors into an object and have it perform the calculations for you.
The elegance of map and reduce (fold) is the essence of functional programming. How do you model map and reduce with REST? You can't! Polymorphism and encapsulation is the essence of object-oriented programming. Where does it go when everything is a resource? It disappears!
I have worked on projects where the goal has been to design every little part of the system as a free standing module with its own life and versioning, that can be switched in and out, but the artifacts have mostly been deployed together and have rarely given any extra value standing on their own. But they have given us a lot of grief when we tried to build a DRY system.
So, a fat model is the way to go. How fat? As fat as possible, but no fatter. How fat is that? As always this is a judgment call but, err on the side of fatter.
If, by luck or skill, my fat system reaches a workload where it will have to be split over multiple processors or machines, it will not be very difficult to split the system, since the system will be well factored, cohesive and DRY!
Note: "Fat is better" is somewhat related to worse is better by Dick Gabriel