Updating The XPages JEE Support Project To Jakarta EE 9, A Travelogue

Dec 14, 2021, 4:41 PM

  1. Updating The XPages JEE Support Project To Jakarta EE 9, A Travelogue
  2. JSP and MVC Support in the XPages JEE Project
  3. Migrating a Large XPages App to Jakarta EE 9
  4. XPages Jakarta EE Support 2.2.0
  5. DQL, QueryResultsProcessor, and JNoSQL
  6. Implementing a Basic JNoSQL Driver for Domino
  7. Video Series On The XPages Jakarta EE Project
  8. JSF in the XPages Jakarta EE Support Project
  9. So Why Jakarta?
  10. XPages Jakarta EE 2.5.0 And The Looming Java-Version Wall
  11. Adding Concurrency to the XPages Jakarta EE Support Project
  12. Adding Transactions to the XPages Jakarta EE Support Project
  13. XPages Jakarta EE 2.9.0 and Next Steps
  14. XPages JEE 2.11.0 and the Javadoc Provider
  15. The Loose Roadmap for XPages Jakarta EE Support
  16. XPages JEE 2.12.0: JNoSQL Views and PrimeFaces Support
  17. XPages JEE 2.13.0
  18. XPages JEE 2.14.0
  19. XPages JEE 2.15.0 and Plans for JEE 10 and 11

I think it's been a little while since I talked about the XPages Jakarta EE Support project of mine. The goal of that is sort of the inverse of the XPages Runtime project: rather than bringing XPages to a proper modern app server, the JEE Support project brings a handful of current Jakarta EE specs to XPages. It started out a few years ago as a sort of proof-of-concept, but I've since been using it for client work to do things like use newer Jakarta REST (n?e JAX-RS), CDI, and newer EL in XPages and OSGi bundles.

The Specification Move

Originally, I targeted a set of specifications from Java/Jakarta EE 8. Some of these were new to Domino outright, while some (such as JAX-RS) existed in the XPages stack already but in very old forms. I implemented those and for a good while just used the project as a source of parts for client work, tweaking it here and there as needed.

However, the long-prophesized package-name switch from javax.* to jakarta.* came to fruition in Jakarta EE 9, released a bit over a year ago. In the intervening year, most implementations of the specs made the switch, and the versions I was using started to show their age (for example, I was using RESTEasy 3, which was already old when I adopted it, and it's going to 6 now). Beyond just the philosophical sadness of my project being behind, I started to grow specific needs to upgrade components: we switched to JSON-B a while ago, but then some new bug fixes in Yasson were coming only to post-jakarta.* builds.

The Initial Work

I first gave a shot to this in August, initially planning to move only JSON-P and JSON-B over to the new namespace. However, I quickly hit the limits of that, since a lot of these specs are interdependent. JAX-RS using JSON-P and JSON-B to emit JSON content, Yasson has some ties to CDI, and so forth. I realized that it was going to have to be all-or-nothing.

So I rolled up my sleeves and assessed the task ahead of me. At a basic level, there was the job of updating my dependencies, which immediately had some good aspects and bad aspects:

  • Previously, I was using a hodgepodge of spec packages like the JBoss bundling of JAX-RS in order to get something that would work and be license-friendly. Now that it was all over at Eclipse, I could switch to the nice, clean official versions and have no license worries.
  • I also used to have all sorts of OSGi rule overrides to account for Domino-specific oddities like ancient versions of various specs being supplied by the default classpath or other, conflicting bundles, all with no versioning. Once I was looking for e.g. jakarta.annotation instead of javax.annotation, I was no longer bound to that particular nightmare.
  • Not all of my dependencies were ready. When I first started, RESTEasy (my JAX-RS provider of choice) had not yet uploaded a JEE-9-compatible version. My main choices were to try using Eclipse Transformer, which would add a whole new layer to the task, or to switch to another provider.

Then there's the elephant in the room: the freaking Servlet API, which much of this depends on. Since the Servlet API is the job of the web container, I can't realistically upgrade it. Fortunately, that's only half true: I can't give it new capabilities (like Web Sockets), but I can wrap the old stuff with the new. And, like the other specs, the switch of the package name was a tremendous blessing, allowing me to deploy the official Servlet 5 API unchanged. Then, I did the tedious work of writing a slew of adapter classes, half wrapping a javax.servlet component and pretending it's jakarta.servlet and half going the other direction. Since the methods are either direct analogs, optional features, or can be emulated, this was actually much easier than I thought it would be. And there: Servlet 5 on Domino! Kind of!

The Showstopper

However, I soon hit what seemed to be a show-stopper: a LinkageError problem when using CDI that didn't show up previously. My search for the topic found only one hit: an issue in Open Liberty referencing almost exactly the same problem. My heart sank when I read that their fix was to upgrade the Equinox runtime - something that's outside my powers on Domino (probably).

So, disheartened, I set it aside for a couple months. I figured there was a small chance that Weld (the CDI implementation at the heart of the trouble) would put out an update that fixed it - after all, an older version worked.

Resuming Work

After setting it aside, it kept eating away at the back of my mind, and two things kept pushing me to go back to it:

  • I'll need to do it eventually. I (and my client projects) can't just be stuck at the old style forever.
  • I really didn't want to admit defeat and switch back to Gson for JSON processing.

So I went back to it. My initial hope - that a new version of Weld would magically fix the problem - proved to not have come to fruition. Still, though, I wasn't sure that it was the exact same problem Liberty encountered. For one, my use of CDI studiously avoids actually telling it about OSGi, since I've had little luck making use of that with Domino's OSGi stack. That was enough cause to make me think I could work around it.

And work around it I did! The trouble turned out to be, unsurprisingly, a bit esoteric, but boiled down to the runtime re-registering proxy classes for the same core components. My guess is that, somewhere along the line, Weld changed some sort of internal cache in a way that would break when using a bunch of ephemeral per-NSF containers as I do. I implemented my own (since it's an intended extension point) and added a bit of a cache, and I was back to the races.

As a convenient blessing, RESTEasy released 6.0.0.Beta1 just days before I got back to it, a major release targeted at JEE 9. That meant that I could save a ton of work by not having to re-work everything for another JAX-RS implementation. I had been looking into Jersey, which I'm sure would have done the job, but it's fiddly work trying to put all these pieces together on Domino, and I was all the happier to not have to re-do it all.


But then I hit a new problem: the javax.mail API, now jakarta.mail. The first part of this is easy enough: bring in the new spec bundle and everything will point to it. Great! I hit an immediate problem, and one I had been dreading dealing with. Though the spec changed package names, the implementation didn't. That brought me face-to-face again with an old nemesis of mine, sitting there in Domino's classpath, corrupting it:

A screenshot of Domino's ndext directory

The way the Mail API works is that there's a file, called "mailcap", that lists implementations for common data types, like:

text/plain;;		x-java-content-handler=com.sun.mail.handlers.text_plain
text/html;;		x-java-content-handler=com.sun.mail.handlers.text_html
text/xml;;		x-java-content-handler=com.sun.mail.handlers.text_xml
multipart/*;;		x-java-content-handler=com.sun.mail.handlers.multipart_mixed; x-java-fallback-entry=true

So, while all the entrypoint classes are jakarta.mail.* now, the implementations remain com.sun.mail.*, all with the same class names. And, since this little jerk of a JAR is sitting in the system classpath, it has a way of showing up all the time, complaining that com.sun.mail.handlers.text_plain is incompatible with jakarta.activation.DataContentHandler.

This is extremely fiddly to deal with, potentially involving writing a special ClassLoader implementation that blocks calls down to the lower-level JAR. While maybe possible, I'm not sure it'd be possible in a way that would be practical for normal use in an app.

And so, with a heavy heart, I forked the thing and added an "org.openntf" in front of all the package names. And that... works! It works just fine. It means that I'm on the hook for manually integrating any upstream changes, but at least it works without having to worry about any conflicts.

That wasn't the end of my trouble with this spec, though. The spec package itself, in jakarta.mail.Session uses ServiceLoader to look for services, and it uses it in the form that looks them up with the current thread's ClassLoader. Because I'm working in OSGi, that ClassLoader - the XPage app's loader - won't know about the implementation classes directly, and this call fails. And, while there's a whole sub-spec in OSGi for dealing with this, I've never had success actually getting it working in Domino.

So I forked that freaking thing too and modified the calls to use its own ClassLoader, which could find the implementation by way of it being a fragment bundle attached to it.

And, with that, finally, I had Jakarta Mail properly hooked up and working without having to jump through too many hoops. I'd still prefer to not have forked the source, but it was the best of a bad lot of choices.

The Final Tally

That brings the specs updated/added in this project to:

  • Expression Language 4.0
  • Contexts and Dependency Injection 3.0
    • Annotations 2.0
    • Interceptors 2.0
    • Dependency Injection 2.0
  • RESTful Web Services (JAX-RS) 3.0
  • Bean Validation 3.0
  • JSON Processing 2.0
  • JSON Binding 2.0
  • XML Binding 3.0
  • Mail 2.1
    • Activation 2.1

Not too shabby, if I say so myself. Technically, Servlet 5.0 is in there, but it doesn't actually bring any newer-than-2.4 powers to the Servlet container, so it's really just infrastructural details.

Now I'll just have the work of updating my client project and finally getting to use whatever that Yasson bug fix was that prompted this in the first place.

New Comment