David Jencks wrote:
On Sep 15, 2008, at 11:41 AM, Joe Bohn wrote:
David Jencks wrote:
On Sep 15, 2008, at 10:45 AM, Joe Bohn wrote:
David Jencks wrote:
On Sep 15, 2008, at 6:18 AM, Joe Bohn wrote:
Thanks for making these changes David. I think this mostly
addresses my concern that plugingroups can be used anywhere
plugins are used (it certainly helps now that they are all cars
:-) ).
So with the subject change ... is it now true what Lin mentioned
earlier in the "[DISCUSS] plugingroups - another idea" thread:
Lin Sun wrote:
> For option 1, I guess I don't understand why we need the
classLoader
> attribute. I think for plugin groups, we should not add the
plugin
> group itself to the classloader, but we should add the
dependencies to
> the classloader. For instance, if we support Joe's scenario,
when a
> user specifies a plugin group as a dependency in his deployment
plan,
> it would just add the dependencies that the plugin group depends
on to
> the project's classloader.
Will the dependencies of the plugingroup be added to the project's
classloader? That would certainly resolve any concerns.
I _think_ they will be in the maven build's classloader but
definitely not in the geronimo classloader. While this is
inconsistent I don't yet see a reasonable use case for a
plugingroup that doesn't have a classloader but does indicate a
group of plugins that need to be added to some projects
classloader. I needed the new functionality for other purposes so
I haven't spent a lot of time thinking about this but a real
concrete use case would go a long way for me.
The scenario I had in mind is very simple.
1) A user has decided (for whatever reason) that they need a core
server built on Geronimo including Jetty & Axis2.
2) The user will build multiple web applications and combine them on
server images as the need arises and so they don't want to include
the applications themselves in a custom server assembly.
2) To address the core server need, the user chooses to assemble
their own server image with the two universal pre-reqs: the
web-jetty and webservices-axis2 plugingroups.
3) The user then goes about building their web applications and
generating geronimo plugins for these web apps to ease deployment.
They want to ensure these applications are installed in the correct
server configuration so they set the prereqs for the applications to
match the core server image that they had earlier constructed - a
prereq on the web-jetty and webservices-axis2 plugingroups.
4) The user will then install the server in the appropriate
location(s) and deploy combinations of the applications on the as
necessary on the server instances.
Of course it is more correct to have the applications prereq the
individual plugins necessary for jetty and axis2 runtimes rather
than the plugingroups. However, that would mean that the user must
understand the detailed plugins for building applications but only
the plugingroups for building server images ... and the whole point
of the plugingroups was to simplify things for the user so they
could deal with higher level concepts. If that simplification makes
sense for assembling a server (a relatively infrequent activity) why
would it not also make sense for assembling an application (a
somewhat more frequent activity)?
And yes, I understand (as will the user) that the plugingroups may
be pulling in more than they really need to run the applications -
esp. if we continue to combine core function, deployers, and console
extensions in the plugingroups. That may be a valid reason to alter
the granularity of the groups we are creating. If not, another
reason to split things up a little finer is to facilitate the
construction of servers without deployment capabilities - but that's
really a separate discussion.
To me it seems more like you are suggesting plugin groups as a
workaround to another problem in the current plugin system, and I
don't think your suggestion will really help.
If you are using maven to build plugins, currently you have to
specify explicitly in the car-maven-plugin which deployment plugins
you want to use, and include them as (provided) dependencies in the
pom. This means you need to know what you want quite explicitly, and
I don't see how plugin groups can help with this yet.... certainly an
area for discussion.
Now, the deployers all add the required dependencies to the plan,
such as in your example axis2 and jetty, so the app certainly won't
start without those present. However at the moment we have no way
for the c-m-p to find out about those added dependencies and add them
to the geronimo-plugin.xml so you have to list them yourself.
It seems to me that if we can somehow use plugin groups in the c-m-p
configuration to specify more concisely what deployers we want, and
get the dependencies added by the deployers into the
geronimo-plugin.xml, then we will have the convenience you want
without the extra classloader or funny "this isn't really a plugin,
include my dependencies" behavior that I'm nervous about.
Yes, you are correct that the deployers are an issue when configuring
the c-m-p. That is a important issue, but not the one I was trying to
discuss here.
I think they are related.... see below :-)
No doubt they are related. I was only trying to tackle one thing at a
time. If plugingroups were interchangeable with plugins we could
leverage deployer plugingroups in c-m-p configuration as well.
If you look at the jaxws-calculator sample it demonstrates something
similar to my contrived scenario. The jaxws-calculator-jetty plugin
has dependencies on 3 other plugins - jetty6, jasper, and cxf (to
match my scenario above let's just pretend this is axis2 rather than
cxf). If the user is already familiar with a set of plugin groups
they would probably find it more intuitive to declare dependencies on
the groups rather than individual plugins. In this example they could
declare dependencies on web-jetty and webservices-axis2 to match what
they had specified when assembling the server.
What I'd like to see is that the user configures the c-m-p so it works
and deploys against the jetty web container, jasper jsp stuff, and axis2
web services. I'd like everything else to flow from that without any
more user input. What's missing right now is
- easier way to set the required deployers (although it isn't too hard now)
- getting the default environment dependencies from the builders into
the geronimo-plugin.xml
If you are deploying something into a running server I want the
generated geronimo-plugin.xml to include all these default env
dependencies as well: the result should be that the user only has to
specify what server bits they want once, either in the c-m-p "which
deployers to use" configuration or what is running in the server they
deploy against.
To reiterate over and over again once more :-/ I think getting the
default env deps into the geronimo-plugin.xml removes any usefulness for
the plugin groups as explicit dependencies for user applications. I
think we might look for a way to simplify or automate choosing the
deployers to use. I don't know if something like plugin groups might be
useful here, or not. We might look into byte code dependency analysis
like dependency:analyze uses to detect use of servlet, ejb, connector,
jms, .... classes and pick them from some kind of list, but I'm not sure
this relates well to the plugin groups you are thinking of. I'm
thinking of a list that says "jetty-deployer, cxf-deployer,
openejb-deployer, jasper-deployer..."
Ok, I see what you are saying. Rather than call out the dependencies
directly and in addition to that call out the deployers in the c-m-p
configuration - the user could skip the dependencies and rely on the
deployer's default environment to fill them in.
Perhaps that will work most of the time for plugingroups that we create
that also have some related deployer (and maybe all of our plugingroups
will have deployer equivalents). But will all dependencies included in
a plugingroup also be pulled in by some deployer's default environment?
However, what happens with user or 3rd party created plugingroups? Will
they always have comparable deployers and hence default environments so
they don't need to be specified as a dependency?
Perhaps the user wants the base server to include some core Geronimo
bits (Jetty, Axis2, etc...) and some of their own functions ... perhaps
a plugingroup of utilities provided by some 3rd party. For example,
let's say there is a plugingroup that includes several common debug
utilities like configurable logging, first-failure data capture, policy
validation, etc... let's call it "debug utilities". Next the user
assembles a server using a combination of the geronimo plugingroups and
"debug utilities" for a base server image. They plan on using this
image multiple times to manage multiple configurations rather than
creating unique assemblies. If they want to build plugins for their
applications that will be deployed to this image, they can configure the
c-m-p deployers and thereby pull in the default environment dependencies
for the geronimo plugingroups. But they still have to include the
individual plugins from "debug utilities" as dependencies if we don't
support using the plugingroup in as dependencies.
Again, my goal is only to get some consistency for the end user. I
still think it's a good idea that our users will appreciate.
Some more rambling thoughts ....
While I understand and appreciate the idea of configuring the deployers
for c-m-p and having them pull in the necessary default environments, it
seems a bit backward from a user perspective. If we want users to
embrace plugins and using maven to construct them, I think it would make
more sense to a user if we approach the problem from the other angle. A
user creating a plugin probably already has some idea of the
dependencies they will need (either plugins or plugingroups if
appropriate). If we let them specify the dependencies we should be
able to determine which of those dependencies requires a particular
deployer (at least dependencies of ours). We could make it more general
purpose by always naming the deployers consistently based upon the
module that they support (ex. -deployer in the artifactId). I'm not
sure how we would accomplish that (perhaps another maven plugin that
front-ends c-m-p) or how we would know if a deployer exists for a given
plugin. But it seems a bit more intuitive to me than telling the user
figure out the c-m-p deployers based upon their dependencies and then
omit the dependencies so that the deployers can pull them in via the
default environment.
I hope I'm not driving you crazy ... it just seems obvious to me that
users will want to leverage plugingroups the same way that they leverage
plugins.
Joe
thanks
david jencks
thanks
david jencks
Thanks,
Joe
thanks
david jencks
Thanks,
Joe
