On Oct 8, 2008, at 11:04 AM, Jason Warner wrote:
I'm not sure if these steps are reasonable from a purely user perspective. When using plain old tomcat, you can download a binary, add your custom valve jar, make a config change and then use your server with its custom valve. To accomplish the same task in geronimo, we are asking the user to download and install maven as well as grab source code for the tomcat plugin. I'd really like to have a way we can accomplish the same goal while allowing the users to maintain a user level of interaction with geronimo.
I think (1) is really a more realistic approach philosophically so I'll only discuss it more.
Lets consider the results of the modifications on tomcat and geronimo.
In tomcat, the user has modified their server installation and has no built-in record of what they did. If they install another server somewhere else they have to look in their notes or try to remember what they did or ??? to get the same result.
In geronimo + maven they have a reproducible and automated way to generate the customization that is suitable for storing in scm, auditing, running through qa, etc etc.
Its also possible to fish the plan out of the tomcat6 plugin, modify it a bit, and deploy it using gshell or (if you didn't start it) using the console. I think you could add the geronimo-plugin.xml using the admin console and add the artifact-aias. This on export would result in a reusable plugin. I'm not sure if you could turn around and install the plugin on the server it was generated on to install the artifact alias so on the next startup you'd get the new tomcat plugin.
My philosophical objection to adding valves to the existing tomcat config is that you've changed it in a fundamental way so you should have a new, replacement, plugin instead. By this point you can add the extra jar(s) anyway as dependencies.
Maybe we could have a tomcat server portlet that would help with generating tomcat server plans with custom valves and connectors and such stuff. I think that right now that is still the hardest part.
On Wed, Oct 8, 2008 at 1:22 PM, David Jencks <firstname.lastname@example.org>
On Oct 8, 2008, at 7:45 AM, Jason Warner wrote:
Could you describe to me in a little more detail what you were thinking in regards to defining a new tomcat server in a child classloader? I'm still working on creating an example, but I found some documentation confirming tomcat's use of a TCCL in loading components and would like to continue the discussion. It seems you are proposing that a user create a plugin that defines a new tomcat instance that includes their custom valve. Am I understanding correctly? I've taken a look at the app-per-port sample you described and this does not seem like a trivial task.
app-per-port is complicated by the additional features there of:
- only one artifact (an ear) instead of 2 or 3 plugins
- starting the connectors after the web app has started
If neither of these features is needed you can just build a plugin with the tomcat server + custom valve. There are two strategies:
1. replace the tomcat6 plugin
2. use the (stopped) tomcat6 plugin as a parent for the new plugin.
In either case I'd build the new plugin with maven and start by copying the tomcat6 plugin and renaming it appropriately. Then modify the plan to include the custom valve.
for (1), you'd just add the jar with the custom valve as a pom dependency. Use an artifact-alias so your tomcat plugin will replace the usual tomcat6 plugin.
for (2), you'd replace the pom dependencies with a dependency on the tomcat6 plugin, and add the custom valve jar dependency. In the c-m-p configuration you'll want to specify the import on the tomcat^ plugin as "classes" so it wont get started. An artifact alias won't work here so don't deploy things that depend on tomcat6 as that will result in the tomcat6 plugin starting and having port conflicts with your plugin.
Building a custom server including your plugin or installing it on a framework server via gshell is likely to work better than trying to replace the tomcat6 plugin while it's running through the admin console.
hope this helps
On Mon, Oct 6, 2008 at 1:59 PM, David Jencks <email@example.com>
On Oct 6, 2008, at 10:35 AM, Jason Warner wrote:
On Mon, Oct 6, 2008 at 11:56 AM, David Jencks <firstname.lastname@example.org>
On Oct 6, 2008, at 7:22 AM, Jason Warner wrote:
On Fri, Oct 3, 2008 at 6:55 PM, David Jencks <email@example.com>
On Oct 3, 2008, at 12:51 PM, Jason Warner wrote:
Hey all. I'm working on an idea for allowing custom valves to be defined in config.xml. Currently this isn't possible since the tomcat classloader would not contain the custom classes for the valve. I've create a jira for tracking this issue  and it contains a few links to workarounds. IMHO, The solution we should be looking for is a way to add classes to a module without having to undeploy, modify the module config, and redeploying.
People have suggested stuff like this before. IMO it pretty much goes against the fundamental idea of geronimo of having fairly fixed plugins with only a few knobs to turn to adjust things in config.xml and config-substitutions.properties.
Why is changing the classloader contents in config.xml a good idea? What is so hard about redeploying the app if you want to change its classloader significantly? If you want to change a class in the app you have to redeploy it.... why is this situation different?
The specific instance I have in mind for this change is using a custom valve for tomcat, so I think the scope really should be limited to just the tomcat module. I can't think of another instance where this would be useful, so it's probably not necessary or desirable to expand it further. I believe this situation is different because the structure of geronimo is causing a disconnect between the functionality of tomcat and the functionality of tomcat as it is embedded in geronimo. As Don just said in the middle of my typing this, I don't believe we should expect the average user to have to rebuild one of our modules to add something that can be added in a much simpler way within tomcat itself.
Could you explain more about the circumstances for this custom valve? Is it intended to be for every app deployed on this tomcat server instance rather than for one particular app? Will it work if it is in a child classloader of the tomcat plugin classloader?
When a valve is added to the tomcat valve chain, it becomes part of the request processing pipeline. Every request that is made to that tomcat server instance passes through this valve chain as it's processed regardless of whether the valve will act upon it or not. It's possible that a single web app will be the only app to use the valve, and for that instance it is already possible to define the valve in the context of the web app rather than the tomcat server. We need to be able to define a valve as part of tomcat server instance as well, though, to be consistent with tomcat. Currently we can only define the valves on the per web app basis.
I don't think this would work in a child classloader of the tomcat plugin classloader. When we start up the tomcat module now, the currently defined valves are processed and added to the engine. The custom valves would need to be added to the valves already in the tomcat engine to be available in the way described previously. Once the valves were added to the engine (which would be using the tomcat classloader, I believe) the class def not found issues we currently see would pop back up. For this to work, the custom valve classes and the tomcat engine would need to share the same classloader.
Could you try this to be sure? I would hope that tomcat would use a TCCL or supplied classloader for loading components rather than something like TomcatEngine.class.getClassLoader() which I believe is what you are suggesting it does.
One example of an inconvenient tomcat configuration is the app-per-port sample where we set up a whole additional tomcat server in a child configuration. I think all the server components in that example are also in a standard tomcat server but its a similar situation to what I'm thinking of here in terms of configuring a tomcat server in a child classloader.
Sure. It'll take me a bit as I don't actually have any examples prepared yet.
At the moment I would MUCH rather see us make it easier for users to deploy new/different/modified tomcat servers (and other plugins) than introduce a hack to modify classloaders of existing plugins. Our customization story is already too complicated, IMO we don't need to glue on more bits that don't actually fit well.
IMO the best end result for users is to have a new tomcat plugin with the needed extra jars and valve configuration. Lets look for a way to make it really easy for our users to get there.
I agree that a whole new plugin with all desired functionality included would be best for users. Any ideas how to make this easier than it currently is? Perhaps the attribute idea mentioned by Joe could serve as a temporary solution until we can come up with something better.
How would you deal with this in an osgi or spring environment?
If anyone knows how osgi deals with situations like this I'd find it really helpful in considering alternative directions.
I think this can be done by allowing a user to indicate jars that should be loaded by a module within the config.xml. These jars can then be added to the module's classloader for use by the module. I'm not extremely familiar with how our classloader works, but I've taken a look through the code and I think the ability to add to the classloader can be implemented without too much difficulty. I'm not quite sure what type of scope to give this change, though. Should I leave it as a change aimed solely at tomcat valves or should it be expanded to encompass any configuration? I realize this is only a rough idea of what i plan to do, but I'm still working out the details of how to proceed. I'm hoping for some feedback on what I intend to do and possibly some alternate ideas if anyone has some.< br>