[forge-dev] shading!?
Lincoln Baxter, III
lincolnbaxter at gmail.com
Wed Apr 20 13:07:19 EDT 2011
David,
If we use a blank/empty modules descriptor, will JARs in plugin.jar:/lib/*
be loaded and isolated from the rest of the system? If so, that might be
exactly what we need, if combined with an appopriate maven packaging
strategy at build time.
~Lincoln
On Wed, Apr 20, 2011 at 12:34 PM, Lincoln Baxter, III <
lincolnbaxter at gmail.com> wrote:
>
>
> I think what JBoss Modules can give you is a functional, safe
>> multiple-delegate class loader which you can't get with URL class loaders.
>> Specifically, f you want to have a _graph_ of class loaders (as opposed to
>> individual flat, merged classloaders) then you need something like Modules.
>>
>
> I think I just embarrassed myself ;) But can weld even handle this kind of
> separate class-loader situation?
>
>
>> As far as Maven integration, you could write a ModuleLoader which does
>> this for you. You'd have to do version management on your own though.
>> Perhaps you could use Aether for the purpose of resolving and fetching
>> artifacts, and then use information exposed therefrom to construct the
>> module graph.
>>
>
> Right, which is something Forge can already do pretty handily (using
> Aether):
> https://github.com/seam/forge/blob/master/shell-api/src/main/java/org/jboss/seam/forge/project/dependencies/DependencyResolverProvider.java
>
>
>> http://www.sonatype.com/people/2010/08/introducing-aether/
>>
>> Your application does not have to be modular to use JBoss Modules. Though
>> of course in my opinion that makes it much nicer. :-)
>>
>>
>> On 04/20/2011 10:56 AM, Lincoln Baxter, III wrote:
>>
>>> I agree with both of you guys, which is why I am soliciting ideas for
>>> this now.
>>>
>>> I can also say that we do not have one big Classpath right now (unless I
>>> am horribly mistaken.) Forge already supports classpath isolation of
>>> individual plugins, so we're already half-way there. If my understanding
>>> of classloaders is at all right, It would not be too much work to
>>> implement a maven resolver and additional level of Composite
>>> classloaders to achieve this goal.
>>>
>>> I don't believe we need to pull in a heavyweight module system like
>>> OSGI, or force developers to do even the simple configuration for
>>> something like JBoss modules (which afaik, does not have a way of
>>> packaging the actual dependencies with the module, or calling out to
>>> Maven for dependencies - which forge can already do.)
>>>
>>> *This is what we have right now:* (see today.png)
>>>
>>> Notice how the Plugin ClassLoaders may reference Forge APIs, but the
>>> plugins may not access each other's classloaders because there is no
>>> link from Forge APIs to the Plugins.
>>>
>>> *And this is all I think it should take to do what I am suggesting:*
>>> (see tomorrow.png)
>>>
>>> If it is technically possible to achieve with JBoss modules, via adding
>>> in parsing the POM and including dependencies automatically, then I am
>>> all for it, but from what I understand, this would not really be a 1-1
>>> mapping and will not be trivial. I know my understanding of ClassLoaders
>>> may be a bit Naive, but I don't think that we need to use JBoss modules
>>> (or OSGi) to achieve this goal of automatic downloading and loading of
>>> plugin dependency JARs.
>>>
>>> We certainly do not need the ability (yet) to drop an individual module
>>> while Forge is running, because Weld does not support this kind of
>>> disruption in the BeanStore. If you remove a JAR while Weld is running,
>>> you get a big fiery explosion. Same for loading a single JAR, you have
>>> to restart Weld because it will not be aware of the changes, and there
>>> is not a good API to make it so. It still has to build the bean graph
>>> all over again. That is one of the big reasons to use a fully-fledged
>>> module system, and we don't have that requirement. Forge simply dumps
>>> all classloaders (until JDK 7 we can't close them easily, yeah memory
>>> leaks, whatever, not really important right now,) then builds a new
>>> network of ClassLoaders and starts over.
>>>
>>> However, if we can extend JBoss modules with Maven support in this way,
>>> and it's something that David would be interested in, then I think
>>> that's one big reason to do it that way. Until then, if we can do it
>>> quickly my way, I'd like to get it working sooner rather than later.
>>> Please please tell me (I'm sure I don't even have to ask) if there are
>>> gaping technical holes in what I am proposing, but if there are not,
>>> then I'm going to stick to this position until someone bursts my bubble.
>>>
>>> ~Lincoln
>>>
>>>
>>> On Wed, Apr 20, 2011 at 9:33 AM, Paul Bakker <paul.bakker.nl
>>> <http://paul.bakker.nl>@gmail.com <http://gmail.com>> wrote:
>>>
>>> Yeah, I'm afraid that's right. I think the module system integration
>>> should get very high priority before we reach a 1.0 version. It's
>>> something that affects the whole way the core is built up, and it
>>> will be very difficult to change that later on.
>>>
>>> Paul
>>>
>>>
>>>
>>>
>>> On Wed, Apr 20, 2011 at 3:17 PM, Max Rydahl Andersen
>>> <max.andersen at redhat.com <mailto:max.andersen at redhat.com>> wrote:
>>>
>>> +1
>>>
>>> I agree with what you want to do Lincoln, but don't see how that
>>> can be made to happen without a module system.
>>>
>>> Until then everything just will have to live in one big classpath.
>>>
>>> /max
>>>
>>> On Apr 20, 2011, at 14:24, Paul Bakker wrote:
>>>
>>> > And that's why a module system is needed. If people can just
>>> add dependencies there will be duplicate (possibly multiple
>>> versions) dependencies. Each plugin should be in a separate
>>> classloader, and each dependency should be too so that a plugin
>>> never breaks other plugins or Forge itself. It would be a big
>>> limitation if plugins can't use libraries otherwise.
>>> >
>>> > Paul
>>> >
>>> > On Tue, Apr 19, 2011 at 6:14 PM, Lincoln Baxter, III
>>> <lincolnbaxter at gmail.com <mailto:lincolnbaxter at gmail.com>> wrote:
>>> > That's right, but what I'm saying is that I don't want
>>> developers to be responsible for anything but *their code* -- if
>>> they have dependencies, those depenencies will be fetched for
>>> them (or somehow bundled in the JAR file itself, which is
>>> certainly possible, however not my preference.)
>>> >
>>> > If you could write a plugin, reference dependencies in your
>>> POM, and have everything *Just Work* don't you think that would
>>> be a friendlier experience?
>>> >
>>> > ~Lincoln
>>> >
>>> >
>>> > On Mon, Apr 18, 2011 at 5:55 PM, Max Andersen
>>> <manderse at redhat.com <mailto:manderse at redhat.com>> wrote:
>>> > But if your plugin uses multiple jars it is not one jar.
>>> >
>>> >
>>> > /max (sent from my phone)
>>> >
>>> >
>>> > On 18/04/2011, at 19.06, "Lincoln Baxter, III"
>>> <lincolnbaxter at gmail.com <mailto:lincolnbaxter at gmail.com>> wrote:
>>> >
>>> >> That doesn't solve the problem of having to drop jar files
>>> onto the classpath in order for plugins to work. I want one JAR
>>> per plugin.
>>> >>
>>> >> ~Lincoln
>>> >>
>>> >> On Mon, Apr 18, 2011 at 12:59 PM, Max Rydahl Andersen
>>> <max.andersen at redhat.com <mailto:max.andersen at redhat.com>> wrote:
>>> >>
>>> >> On Apr 18, 2011, at 18:57, Lincoln Baxter, III wrote:
>>> >>
>>> >> > I don't want to force plugin-developers to create modules
>>> for every dependency that their plugin requires. That's why I've
>>> been avoiding OSGI or JBoss Modules.
>>> >>
>>> >> But then you shouldn't be forcing them to shade either - you
>>> should just have one global classloader for the plugins then.
>>> >>
>>> >> /max
>>> >>
>>> >> >
>>> >> > ~Lincoln
>>> >> >
>>> >> > On Mon, Apr 18, 2011 at 12:37 PM, Max Rydahl Andersen
>>> <max.andersen at redhat.com <mailto:max.andersen at redhat.com>> wrote:
>>> >> >
>>> >> > > What do you think about using Maven APIs to inspect the
>>> POM and fetch dependencies dynamically for each plugin, then
>>> isolate them in the plugin's classloader?
>>> >> >
>>> >> > Why not just load them in to one classloader so you don't
>>> have collisions when there are mixed dependencies on Forge it self
>>> ?
>>> >> >
>>> >> > How about shared data instances ? How does that work ?
>>> >> >
>>> >> > ...as a side note...creating our own module system now - I
>>> feel that is a very bad direction :(
>>> >> > Might as well adopt osgi plugin system if you want this
>>> kind of separation ?
>>> >> >
>>> >> > /max
>>> >> >
>>> >> > >
>>> >> > > ~Lincoln
>>> >> > >
>>> >> > > On Mon, Apr 18, 2011 at 11:56 AM, Max Rydahl Andersen
>>> <max.andersen at redhat.com <mailto:max.andersen at redhat.com>> wrote:
>>> >> > >
>>> >> > > On Apr 18, 2011, at 15:15, Lincoln Baxter, III wrote:
>>> >> > >
>>> >> > > > "if there is a standard location for dependencies"
>>> >> > > >
>>> >> > > > What do you mean?
>>> >> > >
>>> >> > > Your "standard" for shading is that you put all classes
>>> into the plugin.jar.
>>> >> > >
>>> >> > > A "standard" for dependencies for a plugin.jar could be
>>> "next to the plugin.jar".
>>> >> > >
>>> >> > > Would still have the problem of overlapping jars but
>>> then at least its easier to see where the duplication is.
>>> >> > >
>>> >> > > /max
>>> >> > >
>>> >> > > >
>>> >> > > > Thx,
>>> >> > > > ~Lincoln
>>> >> > > >
>>> >> > > > On Mon, Apr 18, 2011 at 6:20 AM, Max Rydahl Andersen
>>> <max.andersen at redhat.com <mailto:max.andersen at redhat.com>> wrote:
>>> >> > > >
>>> >> > > > > I was thinking we might already be able to do that
>>> using the existing pom.xml metadata that's stored in the
>>> artifact itself, or is that too tricky?
>>> >> > > >
>>> >> > > > if there is a standard location for dependencies then
>>> it should be fine - at least better than shading ;)
>>> >> > > >
>>> >> > > > /max
>>> >> > > >
>>> >> > > > >
>>> >> > > > > On Mon, Apr 11, 2011 at 6:51 PM, Max Andersen
>>> <manderse at redhat.com <mailto:manderse at redhat.com>> wrote:
>>> >> > > > > I was thinking Plugin jar having references to
>>> dependent jars via manifest.mf
>>> >> > > > >
>>> >> > > > > /max (sent from my phone)
>>> >> > > > >
>>> >> > > > >
>>> >> > > > > On 12/04/2011, at 00.39, "Lincoln Baxter, III"
>>> <lincolnbaxter at gmail.com <mailto:lincolnbaxter at gmail.com>> wrote:
>>> >> > > > >
>>> >> > > > >> Can you give an example of how you would bundle the
>>> JARs? (Just put them in /META-INF/dependencies/ ... ?) And would
>>> that not cause just as many class conflicts? If you
>>> shade/relocate then the deps *should be* completely isolated.
>>> >> > > > >>
>>> >> > > > >> On Mon, Apr 11, 2011 at 4:11 PM, Max Rydahl
>>> Andersen <max.andersen at redhat.com
>>> <mailto:max.andersen at redhat.com>> wrote:
>>> >> > > > >> well, recommending just bundling jars would be a
>>> better approach than shading IMO.
>>> >> > > > >>
>>> >> > > > >> /max
>>> >> > > > >>
>>> >> > > > >> On Apr 11, 2011, at 16:00, Lincoln Baxter, III wrote:
>>> >> > > > >>
>>> >> > > > >> > Yeah, shading is currently the recommended
>>> approach. Conflicts should be avoided by using relocations. I
>>> know this is... not a great method, but for now it's all we've
>>> got. Open to suggestions.
>>> >> > > > >> >
>>> >> > > > >> > ~Lincoln
>>> >> > > > >> >
>>> >> > > > >> > On Mon, Apr 11, 2011 at 3:41 AM, Max Rydahl
>>> Andersen <max.andersen at redhat.com
>>> <mailto:max.andersen at redhat.com>> wrote:
>>> >> > > > >> > Heya,
>>> >> > > > >> >
>>> >> > > > >> > Lincoln, I just saw your commits to hibernattools
>>> plugin at
>>> (
>>> https://github.com/forge/plugin-hibernate-tools/commit/8b208b4a8e79dbb8a01d10d266ee81afd2cf7106
>>> )
>>> >> > > > >> >
>>> >> > > > >> > Is shading of jars really the recommended
>>> approach for plugins in Forge ?
>>> >> > > > >> >
>>> >> > > > >> > How are you going to share/avoid collisions of
>>> libraries across plugins if they need to bundle via shading ?
>>> >> > > > >> >
>>> >> > > > >> > /max
>>> >> > > > >> > http://about.me/maxandersen
>>> >> > > > >> >
>>> >> > > > >> >
>>> >> > > > >> >
>>> >> > > > >> >
>>> >> > > > >> > _______________________________________________
>>> >> > > > >> > forge-dev mailing list
>>> >> > > > >> > forge-dev at lists.jboss.org
>>> <mailto:forge-dev at lists.jboss.org>
>>>
>>> >> > > > >> > https://lists.jboss.org/mailman/listinfo/forge-dev
>>> >> > > > >> >
>>> >> > > > >> >
>>> >> > > > >> >
>>> >> > > > >> > --
>>> >> > > > >> > Lincoln Baxter, III
>>> >> > > > >> > http://ocpsoft.com
>>> >> > > > >> > http://scrumshark.com
>>> >> > > > >> > "Keep it Simple"
>>> >> > > > >>
>>> >> > > > >> /max
>>> >> > > > >> http://about.me/maxandersen
>>> >> > > > >>
>>> >> > > > >>
>>> >> > > > >>
>>> >> > > > >>
>>> >> > > > >>
>>> >> > > > >>
>>> >> > > > >> --
>>> >> > > > >> Lincoln Baxter, III
>>> >> > > > >> http://ocpsoft.com
>>> >> > > > >> http://scrumshark.com
>>> >> > > > >> "Keep it Simple"
>>> >> > > > >
>>> >> > > > >
>>> >> > > > >
>>> >> > > > > --
>>> >> > > > > Lincoln Baxter, III
>>> >> > > > > http://ocpsoft.com
>>> >> > > > > http://scrumshark.com
>>> >> > > > > "Keep it Simple"
>>> >> > > >
>>> >> > > > /max
>>> >> > > > http://about.me/maxandersen
>>> >> > > >
>>> >> > > >
>>> >> > > >
>>> >> > > >
>>> >> > > >
>>> >> > > >
>>> >> > > > --
>>> >> > > > Lincoln Baxter, III
>>> >> > > > http://ocpsoft.com
>>> >> > > > http://scrumshark.com
>>> >> > > > "Keep it Simple"
>>> >> > >
>>> >> > > /max
>>> >> > > http://about.me/maxandersen
>>> >> > >
>>> >> > >
>>> >> > >
>>> >> > >
>>> >> > >
>>> >> > >
>>> >> > > --
>>> >> > > Lincoln Baxter, III
>>> >> > > http://ocpsoft.com
>>> >> > > http://scrumshark.com
>>> >> > > "Keep it Simple"
>>> >> >
>>> >> > /max
>>> >> > http://about.me/maxandersen
>>> >> >
>>> >> >
>>> >> >
>>> >> >
>>> >> >
>>> >> >
>>> >> > --
>>> >> > Lincoln Baxter, III
>>> >> > http://ocpsoft.com
>>> >> > http://scrumshark.com
>>> >> > "Keep it Simple"
>>> >>
>>> >> /max
>>> >> http://about.me/maxandersen
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >>
>>> >> --
>>> >> Lincoln Baxter, III
>>> >> http://ocpsoft.com
>>> >> http://scrumshark.com
>>> >> "Keep it Simple"
>>> >
>>> >
>>> >
>>> > --
>>> > Lincoln Baxter, III
>>> > http://ocpsoft.com
>>> > http://scrumshark.com
>>> > "Keep it Simple"
>>> >
>>> > _______________________________________________
>>> > forge-dev mailing list
>>> > forge-dev at lists.jboss.org <mailto:forge-dev at lists.jboss.org>
>>>
>>> > https://lists.jboss.org/mailman/listinfo/forge-dev
>>> >
>>> >
>>> > _______________________________________________
>>> > forge-dev mailing list
>>> > forge-dev at lists.jboss.org <mailto:forge-dev at lists.jboss.org>
>>>
>>> > https://lists.jboss.org/mailman/listinfo/forge-dev
>>>
>>> /max
>>> http://about.me/maxandersen
>>>
>>>
>>>
>>>
>>> _______________________________________________
>>> forge-dev mailing list
>>> forge-dev at lists.jboss.org <mailto:forge-dev at lists.jboss.org>
>>>
>>> https://lists.jboss.org/mailman/listinfo/forge-dev
>>>
>>>
>>>
>>> _______________________________________________
>>> forge-dev mailing list
>>> forge-dev at lists.jboss.org <mailto:forge-dev at lists.jboss.org>
>>>
>>> https://lists.jboss.org/mailman/listinfo/forge-dev
>>>
>>>
>>>
>>>
>>> --
>>> Lincoln Baxter, III
>>> http://ocpsoft.com
>>> http://scrumshark.com
>>> "Keep it Simple"
>>>
>>
>>
>> --
>> - DML
>>
>
>
>
> --
> Lincoln Baxter, III
> http://ocpsoft.com
> http://scrumshark.com
> "Keep it Simple"
>
--
Lincoln Baxter, III
http://ocpsoft.com
http://scrumshark.com
"Keep it Simple"
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.jboss.org/pipermail/forge-dev/attachments/20110420/48a1a985/attachment-0001.html
More information about the forge-dev
mailing list