On May 19, 2010, at 3:06 PM, Guillaume Sauthier wrote:
Le 19/05/2010 14:44, Vincent Massol a écrit :
On May 19, 2010, at 2:02 PM, Guillaume Sauthier
wrote:
Hi Vincent
Here are some comments inlined...
Le 18/05/2010 19:24, Vincent Massol a écrit :
Hi devs,
I've explored enough of OSGi to get a handle of how it works and what it would mean
to use it for XWiki as the basis for our Extension Manager (Thomas helped me with this
too).
The executive summary is that I don't think we need to use OSGi right now for the
Extension Manager and we can defer its usage for later.
Let me start by listing what would OSGi bring for XWiki:
1) Known framework with good documentation, books, etc
2) Users would be able to write extensions using OSGi programming model (without using
XWiki's programming model), this also means the ability to reuse OSGi bundles inside
XWiki
As OSGi use the service layer (µservice as we should say now) to
interoperate, everything that is service can talk with another service. It doesn't
matter if your µservice was produced by hand, with iPOJO, Blueprint (any OSGi component
model you want).
yep
So if the XWiki programming model
produces/consume µservices other µservices can be used too, they just have to conform to
the µservice API/Interface.
yep
3)
ClassLoader isolation and thus ability to have several versions of classes
4) OBR (Bundle Repository) for downloading and installing remote jars
OBR is not something that will help to download and install remote
bundles.
With a "naked" gateway, you can always download and install remote bundles
(http://....).
The goal of OBR is to help installling a *set* of bundles with dependencies between
them.
For example, if Bundle A depends on Bundle B, you should have to manually download and
install 2 Bundles.
With OBR, you just "say": "I want A installed" and B will be
installed too.
Yes, that's actually what I meant by "installing remote jars"
(the jars + their deps) but I should have been more explicit :)
Yep, I've understood what you meant when you were comparing with Maven3 Embedder
(making clear that you talked about dependencies) :)
Now here
are the problems found with OSGi so far:
A) Impedance mismatch with the webapp/WAR model (classloader and visibility issues). Two
examples:
- it would mean having a special directory somewhere, where to put the bundles, and this
means not going through the standard WAR deployment process.
- if we want to keep WEB-INF/lib in order to have a standard WAR deployment then it would
mean keeping for ex the core jars in WEB-INF/lib and using OSGi only for extensions.
However doing this is not really possible since it would mean extension (OSGi bundles)
would have Imports on core classes but this won't work since core classes won't be
OSGi bundles
In that case, It's possible to have the webapp launching an OSGi
gateway.
You could have the core classes in the WEB-INF/lib/*.jar (or WEB-INF/classes) and have a
selected set of packages (from the core) exported by the gateway (through the System
Bundle). So your extensions (standard bundles) can use the packages from the core.
Yes that could be done. We'd need to use the
Constants.FRAMEWORK_SYSTEMPACKAGES_EXTRA config property and manually list all packages
(this is a pain!).
Not all packages, only the ones you want to be exported (basically your APIs).
In fact, you'll choose to export the packages that you think could be used by
extensions.
Yes but this is still some hard work since anytime a new API is introduced we'll need
to remember to update that configuration.
I guess an
alternative would be to create a xwiki core bundle that is empty except for
Bundle-Classpath in its manifest, generated by the maven bundle plugin and pointing to
WEB-INF/lib/*.jar.
Hmm, not sure it will work.
Bundle-Classpath refers to some inner jar file theoretically ...
From what I understand from
http://www.aqute.biz/Blog/2007-02-19 it should work I think.
The only thing you cannot do with that case is to
update the core classes (since they are not really in a bundle, but considered as part of
the JDK packages).
Yes but that's acceptable I think and is not our primary need.
You'll need to simply reload your webapp with the new core classes
Yes.
B) In practice it's very hard to use OSGi with only a part of the modules being OSGi
bundles and another part being standard external libraries. Everything must be OSGified
for it to work well.
You're absolutely right, it's easier when everything is OSGified.
In addition we might want not need the features
listed above:
1) Not really needed especially since we have the XWiki programming model that we
don't want to loose right now
2) is a nice to have but not a requirement for the Extension Manager
3) in practice it's going to be very hard to support multiple versions of classes.
Imagine for example an application A which is providing a macro in a wiki page.
Application B requires version 1.0 of A and application C requires version 2.0 of A. How
do we support that? OSGi won't help here since this is about wiki pages. It's only
helping for jars. The jar isolation could also easily be implemented using a custom URL
classloader without too much problem IMO, but I'm inclined to say that we don't
need to support multiple versions in a first version of the Extension Manager.
Agree, OSGi will help you to bind a bundle to an exported package using a
specified version (may be a version range).
If what you produce is a wiki page, not some java code, it will be more difficult to
requires a given package version...
As OSGi targets Java, it's more usable (and there are more tools available) if you
write directly Java code: Bnd can automatically deduce your imports and service-oriented
component models handles dynamic service dependencies for you.
Note: I like what I saw in OSGi and I still think one goal of xwiki should be
to integrate with OSGi (or with the JSR294 to be more generic, if that JSR is implemented
one day... ;)). It's just that for the Extension Manager use case, I'm not sure we
need OSGi from the onstart.
This time, it's me that doesn't really know what will go in this
ExtensionManager, and more generally, what is an Extension :)
All depends of what extensions are.
It's a mix of wiki pages, jars containing xwiki component implementations (like a
xwiki java macro) and external jars (third party libraries).
See
http://dev.xwiki.org/xwiki/bin/view/Design/ExtensionManager for more details of what
we'd like. Note that I need to update this page since Thomas has almost convinced me
that we don't need a super XAR format and that we can keep the XAR format for wiki
page as it is now and not have a single jar/zip by using a special empty artifact made
only of a POM listing what an extension is (ie a list of dependencies on XARs and jars).
If you have a small core and everything else provided
as extensions, putting OSGi from the onstart may be a good choice.
If extensions are java classes (that can be seen as services), using OSGi just for them
also may be a good soft introduction to OSGi.
If extensions are not java code, OSGi is a bad idea :)
We have both (and code can be provided by wiki pages in the form of scripts and macros)
but we cannot have different versions of wiki pages right now inside a single xwiki
instance (I'm not talking about page history but about branches, each having its
history). Another problem I mentioned would be to decide what classloader to use when
executing a {{script}} macro.
Thanks
-Vincent
Thanks for the
info Guillaume,
-Vincent
PS: I'm still discovering all this and I'm not yet 100% sure of what I say.
There's a little nagging thought in the back of my head that tells me that going with
OSGi now is the best thing to do even though we might not need all its power (and
complexity) right now but that it would be easier to shift now than later... ;)
I understand, moving to OSGi for JOnAS has been a long work (and still not completely
finished to some aspects), but we started with a non modular architecture, spaghetti code,
... :'(
In your case, I've understood that you have clean APIs, already have modules, so the
move should be easier :)
--G
>
>> Cheers
>> --Guillaume
>>
>>
>>> 4) We can easily use the Maven 3 Embedder to download and install maven
artifacts into a local repository. This is convenient since we already deploy our
applications/modules to a maven repository. Note that the only "difficult" part
here might be searching into a remote repository; I don't know if a maven api exists
for this (but probably). In the worst of cases we could use the Nexus REST API to provide
this feature for now.
>>>
>>> And here's what I think we should do right now:
>>> i) focus on the 2 features we need for the Extension Manager:
>>> -- ability to install an application containing not only wiki pages but also
jars
>>> -- ability to install an application from a remote repository
>>> ii) then once i) is done focus on the upgrade part of the extension manager
>>>
>>> Implementation note:
>>> - To implement i) we can do this easily I believe using a custom URL
ClassLoader (we already have one btw used in the script macro) that we pass to the
Component Annotation Loader when a new JAR is installed (in order to register the
components in that jar) + set that custom URL ClassLoader as the new thread context
classloader (for code that would use the thread context classloader to load classes).
>>>
>>> WDYT?
>>>
>>> Thanks
>>> -Vincent
>>>
>>> PS: I'd be interested to hear from OSGi experts what they have to say
about this since I'm an OSGi newbie and there might be stuff I have not
seen/realized.
>>> PPS: I'm just realizing that the proposal in this mail is close to
Pomegranate's idea:
http://www.caucho.com/projects/pomegranate/