Hi Gianny,
On Mar 17, 2009, at 4:24 AM, Gianny Damour wrote:
On 13/03/2009, at 6:44 PM, David Jencks wrote:
On Mar 12, 2009, at 10:02 AM, David Jencks wrote:
On Mar 12, 2009, at 3:25 AM, Gianny Damour wrote:
On 12/03/2009, at 4:29 AM, David Jencks wrote:
I think I probably have the most experience with classloading
problems in geronimo and the only real problem that arises is
loading a jar in two different classloaders. This can be
solved by a classloader-per-jar model which offers no
theoretical problems to set up in geronimo but practically would
take a lot of work (and maven projects to build a plugin per
jar). So I think we'll have to see what kind of problems we get
with trying to actually use OSGI.
Hi,
Thinking more about this, I believe we can expedite the
implementation of a classloader-per-jar model. Under the hood of
a MultiParentClassLoader we can replace the current
implementation of find class and resources contracts by an
implementation which delegates to a bunch of URLClassLoaders (one
per jar). These bunch of URLClassLoaders are global classloaders,
i.e. shared across all the configs/MultiParentClassLoaders. The
core challenge is to create them in a hierarchy respecting the
maven dependency declarations. So, we could install the pom of
the dependencies in the repo and lazily parse them when
MultiParentClassLoader are created to build this global and share
tree of URLClassLoaders.
IMO the danger here is that the maven pom may not exist or may be
wrong. OSGI has the same problem in that the vast majority of
released jars don't have osgi manifests. I think I saw a rumor
that spring spent a lot of effort osgi-ifying a lot of commonly
used jars to try to solve this problem.
I also don't know if there are situations in which a small number
of closely related jars need to be loaded in a single classloader,
perhaps because one of the jars is "optional" but if present the
"main" jar needs access to its classes. I think there was an osgi
feature that looked sort of like this.
I just started to work on it and I will post back my findings (i
should be able to complete this over the week-end). Even if we
switch to an OSGi kernel, part of this work may hopefully still
be useful.
Unless you are pretty sure we won't have the kind of problems with
bad community metadata suggested above it might be a good idea to
do this in a sandbox branch?
Thinking about this more I really expect the code will be easy
compared to straightening out the dependencies :-(
As a concrete example, the JACC spec needs the servlet and ejb
specs but in our maven poms we've excluded them as dependencies so
as to make it easier to upgrade the jars independently. While
(especially with maven 3) we can probably put in the dependencies
with version ranges, they aren't there now. Getting the server to
work again may be time consuming. I really hope I'm wrong :-D
Hi David,
I just created a patch
http://issues.apache.org/jira/browse/GERONIMO-4590
which provides a preview of the code changes necessary to move to a
one classloader per jar model. It is not yet good enough to create a
branch as only 22 configurations out of the 27 of the jetty-jee5
assemblies can successfully start (failing config is certainly a
maven dependency problem).
The design is rather simple:
1. when a configuraton classloader is created, configuration
dependencies and classPath entries are used to generate a bunch of
GlobalClassLoaders - one per dependency or one for all the classPath
entries.
2. this set of GlobalClassLoaders are put in a hierarchy according
to maven dependency declarations (note that dependency version
declarations are not honored even if present). Maven dependency
declarations can be augmented by dropping an XML file with maven
like <dependencies> declarations.
3. the classloaders of the parent configuration of the configuration
being created are added to this set of GlobalClassLoaders.
4. this set of GlobalClassLoaders is used under the hood of the
configuration's MultiParentClassLoader to implement find class and
resources contracts.
Four maven dependency declarations had to be supplemented to start
22 configurations. Assuming the same ratio, about 14 dependency
declarations may have to be supplemented to start the jetty-jee5
assembly. This is not too involving.
Could you please have a cursory review of this preview patch and let
me know your thoughts?
I finally got some more time to look at this -- thanks for applying it
to my sandbox framework branch.
There are two things that worry me about this.
1. IIUC whenever you include a jar in a configuration all the
configuration's parents get added as parents to the jar's global
classloader, in this code in MultiParentClassLoader2:
protected void addParents(Set<GlobalClassLoader>
globalClassLoaders) {
for (GlobalClassLoader globalClassLoader :
globalClassLoaders) {
for (ClassLoader parent : parents) {
globalClassLoader.addParent(parent);
}
}
}
I don't think this is acceptable. I think that once we have a
GlobalClassloader set up for a jar it needs to be immutable. With
your code which jar a class is loaded from could depend on which
configurations are started when you try to load the class.
I'm sure I haven't thought this through completely but I really wonder
if adding parents to the global classloaders is necessary. If we
ignore geronimo plugins based on javaee apps for a moment, geronimo
plugins won't have any classes in them, and jars will have maven
dependencies on the appropriate other jars anyway. So I think that in
the normal case adding these parents won't add any missing classes.
Constructing a classloader for a javaee app that can be used as a
parent of some other classloader is a geronimo specific feature anyway
not supported by maven so I think it would be OK to just have people
include such dependencies in the <baseURL>-additional.xml (or the pom,
possibly).
If we drop this add-parents behavior we might need to add classloading
rules to the global classloaders and either specify it directly or
push it down from a plugin config. Again I haven't thought this
through.
2. IIUC the <baseURL>-additional.xml only lets you add to the maven
pom. I think we need to be able to delete stuff too.
As a probably easy-to-fix bug I think that we currently decide if
something is a jar or a plugin in MavenDependencyResolver by whether
there is exactly one URL for the artifact. However a geronimo plugin
with classes inside -- say an ejb jar -- could have exactly one url
but not be a jar.
These are just my first impressions, I'll keep looking.
thanks!
david jencks
Thanks,
Gianny
thanks
david jencks
thanks
david jencks
Thanks,
Gianny
One thing I'd really like actual user data on is how people
actually specify osgi classloading info in real life. I'm very
aware that in theory you are supposed to specify the package
imports and exports for your bundle but I've been told that in
real life everyone with a serious osgi project actually
specifies the jar dependencies they want using require-bundle.
thanks
david jencks
Thanks,
Gianny
On 11/03/2009, at 7:11 PM, Guillaume Nodet wrote:
On Wed, Mar 11, 2009 at 08:57, Gianny Damour <gianny.dam...@optusnet.com.au
> wrote:
Hi,
FWIW, I believe that improving the configuration style to
simplify the means of creating a bunch of objects in the
kernel has more benefits than swapping the classloading infra.
On paper OSGi may appear as superior from a classloading
isolation perspective; however, I believe the current CLing
design is nearly up to par with the OSGi one and that the main
challenge is to properly tune export/import dependency
declarations.
I have to disagree with that. The CLing mechanism is very
different in Geronimo (from what I recall) and OSGi. Geronimo
uses a multi-parent classloader style with some nice features
to be able to hide / never override + parent or self-first
delegation.
OSGi CLind is very different: the first one is that you don't
really have parent classloaders: the classloader for a given
OSGi bundle is calculated wrt to the constraints expressed in
the OSGi manifest using imported packages or required bundles.
Let's take an example:
bundle A needs api packages from bundles B and C
implementation classes from bundle B and C needs something
from bundle D but with different versions
OSGi will be able to handle that because of non tree-like
CLind mechanism: if bundle A is wired to bundle B, it does not
have to see all the requirements from bundle B, and same for
C. Therefore, bundle A can be wired to both B and C without
problems because it will not see bundle D at all (so there's
no conflicts between the two versions of bundle D).
OSGi has a much more powerful CLing mechanism where you can
express lots of different constraints. The drawback is that
establishing the classloader can take a bit of time, so going
to OSGi most certainly leads to a big slowdown at startup
while creating the classloaders.
Also, OSGi does not really play nicely with the usual JEE way
to discover implementations through the MANIFEST/services
entries. That's kinda what we've tried to solve in servicemix
specs, though I'm not sure if that really applies everywhere
because I would imagine the classloaders for EARs are not
really OSGi classloaders ...
I certainly don't want to say OSGi is not the way to go, just
want to make the point that there are benefits but also
drawbacks.
The JAXB approach to turn xml plans to a bunch of objects is
certainly interesting. I believe it is still a technology
limiting decision whereby a lot of custom code will have to be
implemented to support various style (factory methods or beans
et cetera) of configurations. I have been bouncing around this
idea a while back and here it is again. Why do we want to
define a XML language to create a bunch of objects when
scripting can do that for us?
I believe that xbean-spring is still unnecessary noisy when
compared to something like the Spring Bean Builder (http://www.grails.org/Spring+Bean+Builder
).
If there is an interest in a scripting approach, then I can
investigate further.
Thoughts?
Thanks,
Gianny
On 11/03/2009, at 6:54 AM, David Jencks wrote:
So as mentioned below I'm starting to look into the osgi
classloading bit, sort of "from the bottom".
Another approach to many of these issues is perhaps "from the
top", from the point of view of going from a presumably xml
plan to a bunch of objects.
I've long thought that it must be possible to leverage jaxb to
do most of the heavy lifting here. In particular sxc is some
code we can presumably actually extend to do stuff like
constructor dependency injection. So another avenue that
could perhaps be approached in parallel would be to
investigate sxc, jaxb, xbean-spring, xbean-reflect, the
blueprint service schema, and jsr299 requirements and see what
we can come up with.
For instance, it might be possible to have a large part of the
blueprint service functionality in jaxb-enabled objects that
jaxb instantiates from the xml. The "init" method could deal
with feeding the metadata into the blueprint service core.
Maybe we can get sxc to use xbean-reflect to create the objects.
So far this is more or less wild speculation in my head...
but I think it would be a lot of fun to investigate.
thanks
david jencks
On Mar 4, 2009, at 4:56 PM, David Jencks wrote:
Geronimo has been around for a while and despite the many good
features gbeans and the geronimo kernel are not catching on
big time. I think we want to consider taking action now to
avoid ending up being dragged down by supporting a dead
container. Here are a few thoughts.
Actual problems with geronimo:
- gbeans are too restrictive. It's too hard to instantiate
other peoples components as gbeans. GBeans don't support
common patterns like factory methods, factory beans, etc etc,
and require the component to be instantiated directly by the
gbean framework.
- it's too hard to get the classloaders to work. The most
common problem is a class cast exception due to loading the
same jar in two plugins. NoClassDefFound errors from an
optional jar in a child classloader are also really annoying.
Really good things about geronimo I haven't seen elsewhere (at
least in one place):
- gbean dependencies work across plugins. Dependencies are a
unified system, not per-plugin.
- gbean dependencies are resolved in the ancestors of a
plugin, not server wide. This means that you can't make a
partially specified dependency ambiguous by deploying
additional plugins. I consider this an extremely important
feature for predictability.
- plugin dependencies allow assembly of a server from the
explicit dependencies which are normally the same as the maven
dependencies.
Other projects and specs that have stuff we should look into:
maven. Maven has a lot better infrastructure for dealing with
dependency resolution from partial transitive dependency
specification than we do. We should look into using more of
their infrastructure.
osgi. osgi has a lot of similarities to geronimo. The osgi
classloading model is getting a lot of people excited. The
import-bundle idea is pretty much the same as our classloader
model where every jar is a plugin. I don't know if people are
really using the allegedly recommended method of specifying
imports and exports and letting the osgi runtime figure out
where they come from; this seems worth investigating to me.
Also, we get periodic inquiries about when we are going to
support osgi and the was ce folks get even more.
osgi blueprint service (rfc 124) This appears to be a simple
wiring framework for a single plugin. IIUC it uses the osgi
service registry for component dependencies between bundles.
xbean-spring. I'd be reluctant to try to implement a
blueprint service that didn't provide the xbean-spring
capabilities really well
ee6 dependency injection. EE6 is going to have a pretty
sophisticated dependency injection service which we'll need to
support anyway. We should try to figure out how much of the
core we can assemble using it.
Other great stuff we have:
xbean-reflect, xbean-finder, xbean-spring
These ideas have been floating around in my head for a long
time and I've chatted with various people about them
occasionally. While more discussion is certainly needed on
everything here I need to do some implementation to understand
much more. So, what I'm planning to do:
Dave's crazy work plan...
- Try to use the osgi classloader. I think this involves
putting the classloader creation in Configuration into a
service. Configurations will turn into osgi bundles. I'll
put the Kernel in the osgi ServiceRegistry so the
Configuration bundle activator should be able to use it to
resolve cross-plugin dependencies.
- try to figure out how maven dependency resolution fits into
osgi.
- see if eclipse p2 is relevant for provisioning geronimo
repositories
at this point I think geronimo would be running on osgi, still
using gbeans.
- look into relaxing the gbean framework so it is more plugin-
at-a-time rather than gbean-at-a-time
- see how that differs from the blueprint service, ee DI, and
xbean-spring. Try to support all of these at once.
Thoughts? Counter proposals? Anyone interested?
many thanks
david jencks
--
Cheers,
Guillaume Nodet
------------------------
Blog: http://gnodet.blogspot.com/
------------------------
Open Source SOA
http://fusesource.com
