I've been trying to pull together some thoughts about what it might mean
for Geronimo to enable itself for OSGi applications and what needs to be
added to the server beyond just adopting an OSGi classloading model.
That sort of change would be primarily transparent for most existing
applications, but to make the change worthwhile, we'd also want to make
Geronimo into a real OSGi application platform.
So, beyond just having the framework environment, what would be the
require elements? Ok, to start with, most (all?) real OSGi platforms
have some concept of a bundle repository. The bundle repository is
where installed bundles are stored and there is generally some sort of
loading/provisioning strategy associated with the repository that
eliminates the need for an application to manually install and start
each of its dependent bundles. I think the characteristics of how the
Geronimo bundle repository is a discussion topic all of its own, so for
now, I'll just assume this piece is there. As part of server bootstrap,
there will be a configured startup of bundles from the repository that
are necessary to bring up the server. This will be similar to the
module bootstrapping the server already goes through. There will also
need to be a mechanism for adding bundles to the repository, probably
both as a command line tool and via the console.
The Geronimo server will need to provision the framework with an initial
set of services that will be available for installed bundles to use.
Some of these services will interact with other portions of the Geronimo
server, while others are platform-agnostic, but provide important
bundle-management services. Looking through the OSGi compendium
specifications, the following look like a good recommended set:
* EventAdmin service (generalized Event broadcast service). This is
fairly self contained, and we can probably just use the Felix
reference implementation.
* Logging service. This is a standardized OSGi logging API. The
reference implementation is just a circular queue and does not
actually log entries to any persistent storage. The Geronimo OSGi
logging service should be integrated with the general logging
support. The PAX logging service looks like a good starting point
for this. I understand that the Geronimo Blueprint service
implementation is already using this version.
* Config Admin. This is a persistent store for configuration data.
I think this one will be an general expectation for many bundles
that are installed on Geronimo, so we'd need to provide an
instance of this.
* UserAdmin service. This is an interface to an authentication
system associated with a platform. I believe this would be fairly
simple to map to the Geronimo authentication services.
* Declarative services. The ability for bundles to declaratively
publish services to the services registry. We'd need to support
this to allow bundles to be used portably across framework host
environments. This should not require any special integration
with the rest of Geronimo.
* Blueprint services. A more sophisticated component assembly
model. This also should not require any special Geronimo
integration.
* Preferences Service. Allows bundles to persistently store
preference information. This is a bundle-driven capability, which
is a bit different than the config admin service. I'm not sure
how prevalently this is used, so this one might not be a requirement.
Interestingly, this diagram of Karaf architecture has quite a bit in
common with what I've just described once you replace "Spring DM" with
"Blueprint Service". There could be an advantage to leverage prior
experience with this environment here.
One key aspect to all of this is deployment and administration. The
Geronimo server will need to provide the conduit for deploying new
bundles to this environment, as well as administrative function. The
OSGi Enterprise Expert Group (EEG) is working on a specification for
using JMX for managing OSGi environments. The reference implementation
for this specification includes a framework neutral set of MBeans for
tracking installed bundles, registered services, config admin, etc.
These look like a good model to follow and can be the basis for
providing console-like administration capabilities. There may be
additional MBeans we'd like to provide for other services, such as the
Blueprint service.
This is probably a good staring point for the discussions. There are
likely other facilities we'll need to add here, but I think this is
probably a good starting point for the discussions.
Rick
