felix-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Peter Kriens <peter.kri...@aqute.biz>
Subject Re: Fun with OSGI
Date Mon, 15 Aug 2011 05:51:21 GMT
I think you run in the problem that you want to have your cake and eat it too. OSGi is implementing
a modular system and enforces this modularity. As you acknowledge, this gives a lot of advantages.
Many of the problems you find are then related to the problems this modularity causes (class
loaders!) and not having access to the implementation details.

The problem is that if you do not enforce modularity, it will not be modular. Unfortunately,
in the Java world there are many popular hacks that are just not modular. Trying to merge
these worlds can be painful. If you have proper modularity, you should not be able to navigate
in your IDE because an implementation could be substituted. Hiding the implementation maybe
makes it harder to use the special options but it eases migration and makes your code more
robust because it has less special dependencies. Integration testing is less valuable but
unit testing is crucial.

Your pain seems to be that you have to give up control in certain areas. I believe that this
is an aspect of modularity/component programming and not inherently OSGi. It is up to you
to decide if the benefits of modularity are worth the pain ... I think they are.

That said, we're currently working on a specification that will allow the OSGi service API
to be used without the class loader model. See http://code.google.com/p/pojosr/ or http://www.osgi.org/blog/2011/04/osgi-lite.html
for more. This will give you the service model but it will not force you to be modular, class
loader hacks still work.

Kind regards,

	Peter Kriens

P.S. Somehow it feels like you see the "container" as too big. It takes +/-100ms to start
a framework. So if you're testing you do not need to create a MockBundleContext. Anyway, MOST
of your code should not see a BundleContext ever. Did you look at bndtools?

On 9 aug 2011, at 20:01, Teemu Kanstrén wrote:

> Hello all,
> 
> I have used OSGI in a few projects, most recently in a research prototype
> for a sensor network data collection thingy.
> 
> Overall, I think OSGI is not hugely complex and provides some useful
> features. However, overall my feelings are a bit mixed. So I would like to
> ask others, what are your experiences in using OSGI vs other platforms. Some
> more specific experiences from my viewpoint:
> 
> -Automated updates (or support for them) are commonly mentioned as something
> supported by OSGI. I see there is some basic support for this in
> loading/unloading services and bundles in the standard container.
> Additionally, there are things like Apache ACE that are commonly mentioned
> to take it further. But I fail to see how this really helps much, the big
> issue for me comes to transferring state from old services to new ones and
> managing all the dependencies between the elements as  change is rarely
> localized. While I have needed to support updates, I find it is easier to
> just deploy a complete new version and restart the software.
> 
> -Service code is separated by OSGI through different classloaders. For me
> this has been really nice in keeping dependencies from spreading and forcing
> me to think about component boundaries in a more focused way.  But running
> an OSGI container just for this seems a bit heavyweight for me. The
> classloader separation also causes some big issues for me such as sharing
> libraries over services, such as web-services frameworks, where managing
> configuration files across services is just extra hard when the classloaders
> are separated, in addition to the usual OSGI classloader issues. When
> libraries are better supported, such as the Felix HTTP service with Jetty,
> it seems nice but actually is a wrapper that hides the configuration options
> under layers of abstraction (added complexity) that hide the more advanced
> configuration options from me and makes it hard if possible at all to use
> them.
> 
> -Managing the framework and my application becomes more complex as the user
> has to understand the container caches, large number of directories,
> libraries, configuration files, etc. Things like configuring my app to run
> as a unix daemon are much more complex to manage and debug as I am not in
> direct control over the platform. Errors in application startup from remote
> deployments are harder since they are shown mainly in Felix and not in my
> application log files.
> 
> -Using OSGI for me is a form of a local-level service abstraction (SOA). I
> have basically used Java interfaces to define the "interface" of each OSGI
> service. But this eliminates the navigation support my IDE in terms of
> static analysis and adds, what seems to me, unnecessary abstration. This is
> probably my failure in using too many interfaces but it feels to me as a
> "common practice" for OSGI apps. I was also looking for some support to hook
> and trace service interactions, which could justify some of this, but there
> is nothing like this.
> 
> -Integration testing is difficult due to all the wiring required to get the
> overall system running. In simple cases it works OK with my own
> MockBundleContext, and the SOA approach makes the component composition even
> cleaner in this regard. However, in more complex interactions starting the
> whole container becomes a big burden for me.
> 
> Overall, the approach of SOA at local level seems great for the overall
> architecture. But it seems to me currently there are just too many issues
> for me, and for that reason, I would prefer a more simple approach.
> 
> Maybe I am just doing wrong or building wrong types of apps. Any other
> experiences?
> 
> Thanks for your thoughts,
> Teemu


---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe@felix.apache.org
For additional commands, e-mail: users-help@felix.apache.org


Mime
View raw message