It would seem that the fine-grained/coarse-grained thoughts have people divided. Rajini's note (aside from the fact she has a tonne of experience having done most, if not all, of the OSGi work in Tuscany) paints a picture where the two are not mutually exclusive. I don't typically like doing two options because that seems like indecision, but in this case they do appear to complement each other.
Based on what I've seen in this thread, I'm hoping this briefly summarizes the collection of thoughts on how we might proceed: Tuscany Running in OSGi 1. Add bundle manifests to all the Tuscany modules (using maven bundle plugin). This will ensure the most fine-grained decomposition works and therefore coarser-grained distributions will also work. 2. Add a distribution which creates bundles around manageable collections of Tuscany modules aligned with how we see the runtime being extended/subsetted/replaced. Have this build and test on Continuum. 3. Create 'virtual bundles' for third-party libraries to avoid licensing and disk space issues (based on Rajini's suggestions, which I need to better understand). 4. Provide guidance on the wiki on how to avoid breaking OSGi. There's also the suggestion that implementation.osgi relate to Distributed OSGi (RFC 119), which shouldn't be lost. Have I missed anything? It sounds like a staging of 1 & 3 first, followed by 2 would work (and 4 if things keep breaking :-( ). I'd be concerned if we didn't get to 2, and as part of 1&3, we need to make sure these are regularly tested under osgi. 2008/5/1 Mike Edwards <[EMAIL PROTECTED]>: > > Simon Laws wrote: > > > On Thu, May 1, 2008 at 10:03 AM, Rajini Sivaram < > > [EMAIL PROTECTED]> wrote: > > > > > > > On 5/1/08, Jean-Sebastien Delfino <[EMAIL PROTECTED]> wrote: > > > > > > > My 2c: > > > > > > > > +1 to promote OSGi to a first class Tuscany runtime environment > > > > > > > > +1 for an OSGi continuum build (thinking about a build profile that'll > > > > > > > run > > > > > > > the Tuscany itest suite in an OSGi environment, similar to the > profiles > > > > > > > we > > > > > > > have for Web containers for example) > > > > > > > > Here's what I imagined we'd do: > > > > 1. add OSGi entries to each of our JAR manifests > > > > 2. have developers maintain them and pay attention to imports/exports > > > > 3. use the OSGi build to detect API and SPI import/export violations > > > > 4. find the best way to OSGi-enable 3rd party dependency JARs > > > > > > > > I realize that my suggestion [1] is not very popular and most people > on > > > > this list would prefer to come up with bigger bundles grouping several > > > > > > > of > > > > > > > our JARs/modules. I don't think that the 'bigger aggregate bundle' > > > > > > > approach > > > > > > > will work, but I'll be happy to watch people try it :) if they want > to. > > > > > > > > With respect to [4] I find rather funny to see many projects out there > > > > claim OSGi enablement without having OSGified their 3rd party > > > > > > > dependencies. > > > > > > > I wonder how that works, can an OSGi-enabled project really leverage > the > > > > OSGi classloader isolation and versioning capabilities when 99% of the > > > > > > > JARs > > > > > > > it requires are not OSGi bundles? I must be missing something... and I > > > > > > > hope > > > > > > > we can do better in Tuscany with a real end-to-end OSGi enablement > story > > > > > > > :) > > > > > > > -- > > > > Jean-Sebastien > > > > > > > > > > > > > > I agree with Sebastien's suggestions1-4. But I would like to suggest a > > > slight variation to the first suggestion. > > > > > > > > > 1. Use maven-bundle-plugin to introduce OSGi manifest entries into all > > > the Tuscany modules, with auto-generated import/export statements. > > > Modify > > > itest/osgi-tuscany to run these modules under OSGi, with all 3rd party > > > jars > > > installed into OSGi using virtual bundles created on the fly. > > > > > > This step will provide a version of osgi-tuscany tests that is less > prone > > > to > > > breakage than the one we have today. It will also help fix any remaining > > > classloading issues that we have left in Tuscany (and hopefully help > > > in maintaining the classloader isolation). This is not a big piece of > work > > > since this is just bringing together the different pieces that we > already > > > have. I will be happy to contribute the code towards this first step, so > > > others can concentrate on what we really want to achieve in terms of > > > modularity, distribution etc. We can also use this step to explore > > > versioning - in particular about having multiple extensions referring to > > > different versions of 3rd party libraries. This will be very useful for > 4. > > > > > > Suggestions 2-3 are not requirements for OSGi, but these are clearly > cases > > > where OSGi technology can help Tuscany improve modularity. If we want to > > > have explicitly hard-coded import/export statements in the modules to > > > enforce modularity, that can be introduced in step 2. > > > > > > And I would expect that there will be more work in terms of building the > > > distributions suitable for OSGi as well as non-OSGi after 1-4. > > > > > > > > > > > > > > > Thank you... > > > > > > Regards, > > > > > > Rajini > > > > > > > > > > Re. modularity, it strikes me that there is lack of consensus about > whether > > big bundles or small bundles, relative to our maven modules, are > preferable > > but that there is consensus about using OSGi bundles to help us test and > > refine out modularity story. > > > > So maybe we should agree to disregard the question about how big OSGi > > bundles should be relative to our maven modules and concentrate on the > > question about whether our maven modules are correctly factored. Using > OSGi > > as a tool to help us with this of course. > > > > Simon > > > > > I support both Rajini's and Simon's points. We can get the mechanics going > first with the bundles following the existing grain of the modules - and we > can discuss and tinker with the modules as we please later to find the best > arrangements. > > I suspect that the differing opinions about module size reflect different > usecases. Lots of small modules can promote great flexibility in the > construction of runtimes, but at the expense of making it much harder for > users to make sense of the options. Larger modules cut down the options but > make things easier for end users. > > > Yours, Mike. >
