Tuesday, March 3, 2009

Clojure + Terracotta: The Next Steps

Update: I've gotten a *multiple* REPLs running with Terracotta. http://paul.stadig.name/2009/03/clojure-terracotta-we-have-repl.html.

In my last post about Clojure + Terracotta I gave an example of sharing specific references between JVMs through Terracotta. This is what I call the "shared somethings" approach. You specify exactly what you like to share. Another approach is what I call the "shared everything" approach.

Shared Everything

The goal of shared everything is to have multiple VMs working within a single global context through Terracotta, all of your vars and refs would be shared by default, and the canonical test case for this would be to define a function in one VM and have it show up automatically in another VM.

The first task was to move my work into a Terracotta Integration Module (TIM). When using a TIM, in addition to packaging the configuration for reuse, classes can be replaced with clustered versions that will work with Terracotta, without having to fork the original code base.

The second was to replace a couple of classes in the Clojure codebase. The Namespace class uses AtomicReference, which is not supported by Terracotta. There was a minor change necessary in Var, too, because it was using its dvals field as a sentinel value to indicate that the var is not bound. This does not do for Terracotta, because dvals is a ThreadLocal, so I created a NOT_BOUND sentinel field. There were some other changes as well, I'm not going to detail all of the changes, but you get the idea.

At this point I would have hoped that I could run a REPL and possibly even try my canonical test case, but it should be so easy. I have run into two major roadblocks:

  1. *in*, *out*, and *err*. Being I/O streams, *in*, *out*, and *err* obviously cannot be shared through Terracotta. The problem is that they are stored as Vars and interned into the clojure.core namespace. This means that Terracotta will try to share them, because they are part of the shared object graph. I could make clojure.lang.Var.root a transient field (through Terracotta's configuration file), but that would make the roots of all Vars transient, which is not what we want. Instead, what I thought I needed is some kind of TransientVar that could have a different root value (not just bindings) for each JVM. I pursued this a bit using the class replacement of the TIM, and concluded that if that is the route to go, then it should probably be made in the Clojure code (it got messy), or that at least there are some changes to the Clojure code that would ease this. What I settled on (after a suggestion from Rich) is to leave the root bindings for *in*, *out*, and *err* as nil and allow the REPL to bind them. However, the REPL did not bind them for me, so I created my own repl.clj file that binds them and calls clojure.main/repl, and it works! However, this is only a temporary solution. Whether it is creating a TransientVar class, or something else, we need a more permanent solution.
  2. Classes. I am able to connect a single JVM to Terracotta, and run the REPL, but I cannot connect multiple JVMs, nor can I disconnect and reconnect a single JVM. When an instance of Clojure connects to Terracotta, it pulls a compiled function out of the object cache, and then throws a ClassNotFoundException because it cannot find the associated class. I started to pursue modifying the DynamicClassLoader and Compiler to store the compiled classes in the Terracotta object graph, and I still think that this might be the direction to go in. However, I wanted to go ahead and share what I have and get some feedback to see if there are any other solutions.

The code is available at http://github.com/pjstadig/tim-clojure-1.0-snapshot/tree/master. In the "examples" directory I have a "shared-everything" example and a "shared-somethings" example. If you have any trouble running the examples, then let me know. There are some dead ends and some commented code that may not make sense, but my goal was to do a proof-of-concept first and to clean it up once I understand what needs to be done.

Conclusion

We are getting close to a "shared-everything" approach to integrating Clojure and Terracotta. We have some issues to deal with, but we are on our way to making this dream a reality.

2 comments:

Taylor said...
This comment has been removed by the author.
Taylor said...

Hi Paul. Very cool - I worked a bit on how to implement a clustered classloader in Terracotta and blogged about it here: java.think(): Clustered Classloader

Hope that's helpful. There's a working example in the Terracotta Forge:

tim-tclib project