Next Steps With the Open Liberty Runtime

Mar 12, 2021, 4:37 PM

Tags: liberty
  1. Options for the Future of the Domino Open Liberty Runtime
  2. Next Steps With the Open Liberty Runtime
  3. Rapid Progress in Open-Liberty-Runtime Land

About a year and a half ago, I wrote a post musing about my options with the future of the Domino Open Liberty Runtime project. It's been serving me well - I still use it here and with a client CI setup - but it hasn't quite hit its full potential yet.

Its short-term goal was easy enough to accomplish: I wanted a good way to run modern Servlet apps using an active Domino runtime, and that works great. Its long-term goal takes more work, though: becoming the clear best way to do "web stuff" on Domino. There are a lot of definitions for what that might be, and that "on Domino" bit may not even be the way one would want to go about doing it. Still, I think there's potential there.

So, this week, I decided to go back in and see if I could spruce it up a bit.

The Core of Domino Java HTTP

This started with me musing a bit on Twitter about the true lowest-level entrypoint in Domino's HTTP stack is, and where the border between native and Java lies. After overthinking it a bit, I found that the answer was obvious from any stack trace: XspCmdManager.

From what I gather, the native HTTP task (which is much more opaque than the Java part) loads up its JVM, uses the code in xsp.http.bootstrap.jar to initialize the OSGi environment, asks that environment for the bundle, and uses XspCmdManager in there to handle the layering.

That class has a couple public methods, but two are of immediate interest: isXspUrl and service. The isXspUrl is called for each incoming HTTP request. If that returns false, then nHTTP goes about its normal business like it always did; if it returns true, then nHTTP calls service with a bunch of handle parameters and lets the Java runtime take it from there.

That got me to tinkering. Since that class is in an OSGi bundle, you can readily "outrank" it by having another bundle with the same name and a higher version available. Then, since the class and its methods are just called by strings (more or less), you can have other classes with the same names and APIs in place to do whatever you want. And, such as it is, that works well: you can pretty readily inject whatever code you want into the isXspUrl and service methods and have it take over.

However, that doesn't actually buy you much. What I'd really want to do would be to improve on the actual HTTP server - HTTP/2 support, web sockets, all that - and the Java layer only comes into play after nHTTP has received and started interpreting the connection as an HTTP request. You're not given the raw incoming stream. Additionally, there's not actually any real need to override this low level: the HttpService classes you can register via the extension point can choose to handle any incoming URL directly at essentially the same low level as XspCmdManager.

So, while that was fun to poke around with, I don't think there's anything really to be gained there.

Reverse Proxy Improvements

So I went back to an older idea I had kicking around for spawning an all-encompassing reverse proxy. The project has had a lesser version of this for a good while, originally as a WAR file you could add to a Liberty server and then later as a lower-level Liberty feature. However, the way that worked was limited: it would allow you to proxy non-matched requests to a Liberty server to Domino, but didn't do anything to coordinate multiple servers beyond that. Additionally, being a Liberty feature, it limited my future options, such as genericizing the project to work with other app servers.

For my next swing at the problem, I went with Undertow, which is an embeddable Java web server in many ways similar to Jetty, and which is (I gather) the core HTTP part of Wildfly. What made Undertow appealing to me was its modern standards compliance, its relatively-low dependency footprint, and its built-in reverse proxy handler. Additionally, since it's Java, that meant I could embed it in the running JVM without spawning yet another process, hopefully making things all the more reliable.

To go with this, the config DB sprouted some more configuration options:

Reverse proxy config

Along with configuration you explicitly set there and in the individual Liberty server configurations, I have the proxy pick up Domino connection information from names.nsf, allowing it to avoid inconvenient extra environment variables or flags.

And, so far, this has been working splendidly. Undertow's configuration is pretty straightforward, and it wasn't too bad to configure it with prefix matching for the context roots of opted-in apps.

The Next Overall Goal

There's more work to do, beyond just finishing the basic implementation here. I'd really like to get it to a point where you can use this to deploy (at least) WAR-based apps "to Domino" without having to think too much about it, like how you don't have to think about deploying an NSF-based app. It should be thoroughly doable to have the reverse proxy pick up its certificate chain from Domino if desired (especially with the revamped capabilities coming in V12), and some recent changes I made make app deployment noticeably smoother than previously.

Certainly, this sort of project has some inherent limitations compared to nHTTP, but this feels like it's getting a lot closer to a direct upgrade and less like a janky proof-of-concept.

New Comment