Hi Jason,

It is implemented and I will check-in over the week-end.

Here is the design:
1. When a ConfigurationData is loaded from a ConfigurationStore, its dependencies can be altered based on scripts matching the pattern dependencies-(.*).groovy. Here is the script I have been using to perform my integration test:

configurationDataBuilder.configure {
addDependency(groupId: "org.springframework", artifactId: "spring-core", version: "2.0.5", type: "jar")
}

2. When the GBeans of a configuration are loaded, i.e. when a Configuration instance is created, GBeans can be updated based on scripts matching the pattern gbeans-(,*).groovy. Here is my integration test script:

import org.springframework.core.SpringVersion

gbeanDataBuilder.configure {
    addGBean(name: 'name', gbean: SpringVersion) {
    }
}

Scripts are searched in the configuration directory, i.e. in the same folder of the META-INF folder of a configuration. This can be easily changed by implementing a ScriptLocater strategy.

I had to add a groovy dependency to the j2ee-system config which is not ideal as all the configurations will now see the Groovy classes. Ideally, I would like to add another configuration where ConfigurationDataTransformers can be declared and the out-of-the-box GroovyTransformer can be specified. This is problematic as such a configuration needs to be started after j2ee-system and before any other configurations. I could add a dependency to this configuration on the innermost configuration after j2ee-system.

Another approach would be to improve the isolation of configuration classloaders, which should also address classloading problems reported by users and the need to fiddle with hidden-classes declarations. Assuming that I stick to the current configuration approach where I declare a GroovyTransformer in j2ee-system, the improvement I am thinking about is: 1) add a new classloading declaration element, maybe hidden-for- children, where users can specify a pattern a la hidden-classes. 2) The above declarations are used to build a classloader which simply delegates to the configuration classloader and filter out classes matching the hidden-for-children declarations. 3) Children configurations are provided with the above classloader instead of the configuration classloader.

With this thing in place, I will be able to add a groovy dependency to j2ee-system w/o having to thing about the impacts to children configurations as I can hide the groovy classes.

If you want me to check-in "as-is" this scripting stuff and revisit the implementation as soon as I figure out which of the two approaches, i.e. add another config or improve classloading isolation, is the best, then let me know.

Thanks,
Gianny


On 29/10/2008, at 2:21 AM, Jason Warner wrote:

Hi Gianny,

Have you made any progress with this? Are you targeting this for the 2.2 release (whenever that happens to be)?

Thanks!

On Fri, Oct 17, 2008 at 8:11 PM, Gianny Damour <[EMAIL PROTECTED]> wrote:
Hi,

I am proposing the following implementation to start with:

1. In the META-INF folder of a config, we scan for files matching the patterns "dependencies-(.*).groovy" and "extentions-(.*).groovy".

2. We execute the scripts "dependencies-(.*).groovy" which modify dependencies. For instance, such scripts may look like:

configurationData + dependency(groupId: 'group', artifactId: 'artifact', version: '1.1', type: 'jar', importType: importType) -- add the declared dependency configurationData - dependency(groupId: 'group', artifactId: 'artifact', version: '1.0', type: 'jar') -- remove the declared dependency

This gives us the final classloader of the config.

3. We execute the scripts "extentions-(.*).groovy" which update the GBeans, For instance, such scripts may look like:

gBeanBuilder.configure {
   addGBean(BasicNodeInfo) { -- add a GBean
       attribute(name: 'node1')
attribute(extendedJMXConnectorInfo: new BasicExtendedJMXConnectorInfo(...))
       reference(referenceName) {
           pattern('aPattern')
           pattern('aSecondPattern')
       }
   }
   removeGBeansMatching('aPattern')
}

gBeanBuilder.updatedConfigurationData

The implementation of such scripts should be as simple as the modification of config.xml.

The fact that they are collocated with the configuration they modify increases cohesion. It would be neat to have such scripts instead of the native or XStream serializations of config.ser.

Let me know your thoughts?

Thanks,
Gianny



On 16/10/2008, at 11:17 PM, Jason Warner wrote:

While David is more interested in the philosophy, I'd prefer to know a little bit more about your thoughts on implementation. Specifically what do you imagine would be involved in defining this configuration? Would it be as simple as a definition in config.xml?

On Thu, Oct 16, 2008 at 4:08 AM, Gianny Damour <[EMAIL PROTECTED]> wrote:
Hi David,

You are correct: the underpinning philosophy of these approaches is to make it easier to modify pre-canned plugins through extension points.

This may be a good approach to improve further the packaging model of dependencies and services. Let's say that an end-user wants to add a connector to the tomcat6 plugin. Instead of using the admin console or updating his config.xml, he can simply deploy a cumulative plan. This notion of cumulative plan is the key differentiator as users can share their cumulative plans "as-is" - i.e. w/o knowing what the plugin to be modified looks like. To some extent, providing a plan ready to be deployed instead of deployment instructions is better for novices.

I will work on the Groovy builder approach and post back more details as I go.

Thanks,
Gianny


On 16/10/2008, at 10:59 AM, David Jencks wrote:

Hi Gianny,

First, I'd like to make sure I understand the philosophy behind your proposals. IIUC they both involve the idea of making it easy to modify an existing plugin rather than making it easy to replace an existing plugin with a similar one.

Why is this a good idea? My idea has been that we should make it easier to replace a plugin with a similar one than modify an existing one, and then we will have the best of all worlds.

All this being said, I think your ideas are both quite interesting. I'm especially interested in the groovy builder approach.

I'll be fairly unavailable until next week but might keep thinking about this anyway.

thanks!
david jencks
On Oct 15, 2008, at 3:46 AM, Gianny Damour wrote:

On 15/10/2008, at 4:16 AM, David Jencks wrote:

That's one of the main missing bits of functionality. Right now the only way to get the g-p.xml is to use c-m-p or to export the plugin from a server it's been deployed into, or to do something by hand with jar packing and unpacking.

The biggest problem here, in my mind, is that jsr88 only wants you to have one "plan": to deploy something you get to specify the artifact and one "plan". Our deployment system is built around jsr88 so we either have to condense the g-p.xml and plan into one "plan" or abandon jsr88.

At the moment I'm thinking that one satisfactory solution might be to more or less embed the plan into g-p.xml. Perhaps we could avoid duplicating most of the dependency info by adding the <import> element to the dependencies in g-p.xml. I guess we'd expect a more or less empty <environment> element in the plan and fill in the dependencies from the g-p.xml when deploying.

I guess another possibility might be to include the info from g- p.xml in the environment element of the plan.

I've been thinking about this on and off for a long time and don't have any solution I'm entirely happy with so discussion and more ideas are more than welcome :-)

Hi,

Another possible solution would be to allow the extension of a given configuration by other configurations. This could work like the web.xml fragment mechanism of the upcoming servlet specs which allows framework libraries to transparently install Web components to the baseline components defined by the web.xml DD.

When a configuration starts it looks for complementing configurations whose responsibility is to alter the baseline configuration. The identification of complementing configurations could be based on a simple naming convention scheme, e.g. if the base configuration is org/tomcat6//car then all the configurations matching the pattern org/tomcat6-transform-DiscriminatorName//car are identified as complementing configurations.

If there are complementing configurations, then the baseline ConfigurationData could be passed to them for arbitrary transformation, e.g. add, update or remove dependencies. An updated ConfigurationData is passed back and actually loaded by the kernel.

The main drawback of this approach is the added configuration complexity. The main benefits is that it provides application server configuration traceability and a mean to perform very simple changes to a baseline configuration w/o having to redefine in its entirety the configuration to be slightly changed.

In another thread about scripting language integration, I suggested an even simpler approach whereby a script is executed to perform ConfigurationData transformations.

If any of these two options are plausible solutions, then I am happy to move forward with an implementation.

Thanks,
Gianny


thanks
david jencks






--
~Jason Warner




--
~Jason Warner

Reply via email to