Pros and Cons of using REST over JMS (and other technologies)

Question

Hey all,
I am working on a project where we were using JMS initially to send messages between servers. Our front end servers have a RESTful API and use JEE6, with EJB 3.1 entity beans connected to a mysql database and so forth. The back end servers are more like "agents" so to speak.. we send some work for them to do, they do it. They are deployed in GlassFish 3.1 as well, but initially I was using JMS to listen to messages. I learned that JMS onMessage() is not threaded, so in order to facilitate handling of potentially hundreds of messages at once, I had to implement my own threading framework. Basically I used the Executor class. I could have used MDBs, but they are a lot more heavyweight than I needed, as the code within the onMessage was not using any of the container services.
We ran into other issues, such as deploying our app in a distributed architecture in the cloud like EC2 was painful at best. Currently the cloud services we found don't support multi-cast so the nice "discover" feature for clustering JMS and other applications wasn't going to work. For some odd reason there seems to be little info on building out a scalable JEE application in the cloud. Even the EC2 techs, and RackSpace and two others had nobody that understood how to do it.
So in light of this, plus the data we were sending via JMS was a number of different types that had to all be together in a group to be processed.. I started looking at using REST. Java/Jersey (JAX-RS) is so easy to implement and has thus far had wide industry adoption. The fact that our API is already using it on the front end meant I could re-use some of the representations on the back end servers, while a few had to be modified as our public API was not quite needed in full on the back end. Replacing JMS took about a day or so to put the "onmessage" handler into a REST form on the back end servers. Being able to submit an object (via JAXB) from the front servers to the back servers was much nicer to work with than building up a MapMessage object full of Map objects to contain the variety of data elements we needed to send as a group to our back end servers. Since it goes as XML, I am looking at using gzip as well, which should compress it by about 90% or so, making it use much less bandwidth and thus be faster. I don't know how JMS handles large messages. We were using HornetQ server and client.
So I am curious what anyone thinks.. especially anyone that is knowledgeable with JMS and may understand REST as well. What benefits do we lose out on via JMS. Mind you, we were using a single queue and not broadcasting messages.. we wanted to make sure that one and only one end server got the message and handled it.
Thanks..look forward to anyone's thoughts on this.

Answer 1

851827 wrote:
Thank you for the reply. One of the main reasons to switch to REST was JMS is strongly tied to Java. While I believe it can work with other message brokers that other platforms/languages can also use, we didn't want to spend more time researching all those paths. REST is very simple, works very well and is easy to implement in almost any language and platform. Our architecture is basically a front end rest API consumed by clients, and the back end servers are more like worker threads. We apply a set of rules, validations, and such on the front end, then send the work to be done to the back end. We could do it all in one server tier, but we also want to allow other 3rd parties to implement the "worker" server pieces in their own domains with their own language/platform of choice. Now, with this model, they simply provide a URL to send some REST calls to, and send some REST calls back to our servers.well, this sounds like this would be one of those requirements which might make jms not a good fit. as ejp mentioned, message brokers usually have bindings in multiple languages, so jms does not necessarily restrict you from using other languages/platforms as the worker nodes. using a REST based api certainly makes that more simple, though.
As for load balancing, I am not entirely sure how glassfish or JBoss does it. Last time I did anything with scaling, it involved load balancers in front of servers that were session/cookie aware for stateful needs, and could round robin or based on some load factor on each server send requests to appropriate servers in a cluster. If you're saying that JBoss and/or GlassFish no longer need that.. then how is it done? I read up on HornetQ where a request sent to one ip/hornetq server could "discover" other servers in a cluster and balance the load by sending requests to other hornetq servers. I assume this is how the JEE containers are now doing it? The problem with that to me is.. you have one server that is loaded with all incoming traffic and then has to resend it on to other servers in the cluster. With enough load, it seems that the glassfish or jboss server become a load balancer and not doing what they were designed to do.. be a JEE container. I don't recall now if load balancing is in the spec or not..I would think it would not be required to be part of a container though, including session replication and such? Is that part of the spec now?you are confusing many different types of scaling. different layers of the jee stack scale in different ways. you usually scale/load balance at the web layer by putting a load balancer in front of your servers. at the ejb layer, however, you don't necessarily need that. in jboss, the client-side stub for invoking remote ejbs in a cluster will actually include the addresses for all the boxes and do some sort of work distribution itself. so, no given ejb server would be receiving all the incoming load. for jms, again, there are various points of work to consider. you have the message broker itself which is scaled/load balanced in whatever fashion it supports (don't know many details on actual message broker impls). but, for the mdbs themselves, each jee server is pretty independent. each jee server in the cluster will start a pool of mdbs and setup a connection to the relevant queue. then, the incoming messages will be distributed to the various servers and mdbs accordingly. again, no single box will be more loaded than any other.
load balancing/clustering is not part of the jee "spec", but it is one of the many features that a decent jee server will handle for you. the point of jee was to specify patterns for doing work which, if followed, allow the app server to do all the "hard" parts. some of those features are required (transactions, authentication, etc), and some of those features are not (clustering, load-balancing, other robustness features).
I still would think dedicated load balancers, whether physical hardware or virtual software running in a cloud/VM setup would be a better solution for handling load to different tiers?like i said, that depends on the tier. makes sense in some situations, not others. (for one thing, load-balancers tend to be http based, so they don't work so well for non-http protocols.)