Yesterday I read an interesting paper called “RETRO: A RESTful Transaction Mode”. On the good side, I have to say, it is one of the best RESTful models I’ve seen thus far. The authors took special care to satisfy the different REST constraints, unlike many “RESTful” services (e.g. twitter that returns identifier and not URIs). On the downside is I think a distributed transaction model is bad for REST or in other words I don’t see a reason for going through this effort and jumping through all these hoops.
- Service Boundary – RESTful or otherwise is a trust boundary. Atomic transactions require holding locks and holding them on behalf of foreign service is opening a security hole (makes it much easier to do a denial of service attack)
- You cannot assume atomicity between two different entities or resources. Esp. when these resources belong to different businesses.
- Transactions introduce coupling (at least in time)
- Transactions hinder scalability – It isn’t that you can’t scale but it is much harder
For rest it is even worse - Since using hypermedia as the engine of state change means that the hypermedia actually describes the protocol, we clutter the business representations (the representations of real business entities like customer, order etc.) with transactional nitty-gritty as the authors say:
“our model explicitly identifies locks, transactions, owners and conditional representations as explicit, linkable resources. In fact, every significant entity in our model is represented as a resource in order to comply with this constraint.”
This also means the programming the resources themselves will get much more complicated
I think that if you want to reap the benefits of REST you should keep the protocol simple and focus on the business and technical merits you can get not bog it all with needless complexity. It seems to me that RETRO is a good mental exercise to show transactions can be RESTful. I think, however that it is an overkill for RESTful implementations.
RESTful architectures will be better off with BASE (Basically Available, Scalable, Eventually Consistent) and/or ACID2 (Associative, Commutative, Idempotent and Distributed) models –or at least the Saga model (which the authors intend to tackle next) which is a better candidate (IMHO) for achieving distributed consensus.