After almost 2 years into my current project, it hasn’t been a week without a hiccup involving JPA and Hibernate. I’m not claiming that Hibernate or JPA are bad tools, nor would I drop it completely for every project, but let’s face it – sometimes it can be the cause of frameworkitis – as a colleague once brought up.
ORM tools sound like a very good idea from an architectural level. No JDBC mapping code to write, relationship management with cascading and all, a powerful query language, all for free. Well, that’s the problem: it’s not for free, at least not outside the sweet spot.
I’ve found that, for JPA (and classic Hibernate), the sweet spot lies in small web applications NOT using EJBs, nor doing any java-based remoting, such as AMF’s or GWT’s RPC. It might also work for pure java desktop applications, such as ALL.com’s.
When you introduce JPA to the Java EE stack though, things start to crumble. First of all, you can’t cheat by following the “Session per Request” pattern suggested by the Hibernate folks – THERE IS NO GLOBAL REQUEST CONTEXT in an EJB. So you get the EntityManager injected by the container into your EJBs, but when you go over a @Remote interface… Whoops, LazyInitializationException! Same goes for things being used by your java web application, be it JSP, JSF, Servlet, or anything else based.
You can fix that by using only @Local interfaces for intra-EJB communication, but then you loose the nice scalability benefits of EJBs, which are the main reason why I use them (or maybe second to free TX). Let me elaborate a bit on this.
Suppose your OrderService invokes your CustomerService for retrieving billing details. If you use a @Local interface, the CustomerService NEEDS to reside on the same VM as your OrderService. If you used a @Remote interface, it could be at any point in your cluster. Doesn’t make sense? Think of this, in an enterprise, the CRM, the e-Commerce and the Backoffice systems all could share the same CustomerService, thus it would make sense to give it more resources, or host it on faster hardware than the OrderService. Making more sense now?
Now the second problem is client/server communication. If things can go wrong inside the Java EE cluster, think of how wrong they could go when you go over the network. Sending an entity over the wire then receiving it back from the client gets it detached from your persistence context. Your collections, entities and everything related to persistence can go cuckoo over that. I’ve had all sorts of errors on this, from LazyInitializationExceptions to insane “transient object found” on relationships, to not even being able to call em.remove on entities sent by the client. This holds true for RMI, BlazeDS / AMF and GWT RPC, and also over XML with JAXB. I’m quite sure the same would happen over JSON or Message Bindings or any other communication protocol, it’s just the nature of the beast.
Now if all that hasn’t scared you yet, there’s relationships and their impact on performance. Lazy load it – you’ll say. Well, easy if you’re in the sweet spot, but outside… BAM, LazyInitializationException, or no values being sent to the client.
Even with lazy loading, performance suffers due to AOP proxying and the such. I’ve recently reconstructed a query that took 5 minutes to load with Hibernate by using plain JDBC and manual mapping to objects, and the same query runs in 0.6 seconds.
Having to run over the objects generated by Hibernate to remove the proxied collections and variables before sending it over RMI, AMF or GWT RPC can double your processing time, as I experienced when using Gilead, and later on using beanlib in my own custom adapter for AMF.
The only solution to this is having a DTO to every domain object you have, so instead of writing Object / Relational mapping code, you write Domain / DTO mapping code, yay! That’s how my last GWT project ended up.
So here’s my word of warning: Skip JPA unless you’re going to use it in it’s sweet spot. If you’re building a Java web application (be it using Seam, JSPs, JSF, Tiles, Wicket, Struts, or whatever other technology) using Spring,Guice or POJO backend components, then by all means do use it. Everyone else, think twice before jumping in the fire.
On the positive side, I will follow up with some ideas on how to break JPA blues on the next post this weekend.
Disagree? Please leave a comment!
UPDATE: In my current project at the Flemish Government’s Roads Agency we’re using jTransfo written by Joachim Van der Auwera to both translate DTOs to Domain Objects and make sure they’re already associated to the Hibernate context. It works like a charm, and I can totally recommend it.