|RE: [equinox-dev] Re: [p2-dev] who should declare dependencies on ds?|
For what it's worth, the Apache developers dodged this problem by not decoupling the packaging of the service interface with implementation of the service. Instead, they relied on declaring imports for the service interface as well as exporting it. Rather than use the interfaces on the internal classpath of the bundle, the OSGi framework would have the opportunity to wire in those same interfaces as exported from some other bundle.
The drawback to this is that if we never decouple the packaging of a service interface from its implementation, we can never exclusively replace an implementation of the service if we do not have its source (we must replace the service interface also). I don't know if this is a good solution to this particular problem (package the egg and flour premixed, to extend the metaphor), but it will work in some cases.
equinox-dev-bounces@xxxxxxxxxxx [mailto:equinox-dev-bounces@xxxxxxxxxxx] On
Behalf Of Jeff McAffer
Sent: Friday, April 02, 2010 4:58 PM
To: P2 developer discussions
Cc: Equinox development mailing list
Subject: Re: [equinox-dev] Re: [p2-dev] who should declare dependencies on ds?
Good point Jason. I would generalize it even more and say that it is not just DI but the decoupling that comes with services (or the extension registry for that matter). We decouple bundles to they don't depend on specific implementations but then do not have a mechanism for spec'ing that they actually need an implementation. There is the component definition in DS etc but p2 or someone else has to read/understand that. The unique thing about DS is that it is even more removed.
You could say , "hey, the bundle has DS markup so it must need DS" While that is likely true in many cases, it is also possible that the same bundle could be used with and without DS. It may contain other markup for other DI mechanisms. These have to be dealt with at a higher level as you say.
On 2010-04-02, at 4:40 PM, Jason Barkanic wrote:
In general this is a problem with any kind of dependency injection, although in this case nothing is actually being injected, but it is being set up and managed by an outside component.
Do you set up a dependency on the dependency injector? How do you best notify clients that they need the dependency injection framework with your config, or else they'll have to set things up themselves? It's not even different implementations of DS, but you could substitute in Blueprint, or Spring, without changing the API (that is if you don't define API to include one particular set of bundles over another).
This kind of thing is annoying though (I've been a victim). I'm interested to see what solutions present themselves as more and more people move to DI and Services paradigms. I think good error messages can help, since that might have alleviated your 6 week search in the first place, but that is easier said than done. The error message could make suggestions about why a service lookup failed, but it's hard [impossible] to really know.
Phil Wrote: =============================================================================================================
I can appreciate the desire to allow different DS implementations but the bottom line is that DS is going to break any RCP application that uses P2 (there may be other fall out as well). My RCP app uses P2 so I thought that I should download 3.6M5 so that I had time to make comments about the API before the API freeze. When I updated not only did my auto update functionality break, but my build server broke also (PDE build with P2). It took 6 weeks of googling before I figured out that there was this new DS bundle that I not only had to include, but I also had to be responsible for starting.
All is well for me now, but I fear that this change is going to have a big impact when 3.6 releases. At a minimum this needs to be documented probably both in "What's New" in the "Plug-In Development Environment Guide" and also in the 3.6 Plug-in Migration Guide. Getting the rcpupdate example updated (bug 307558) was a good step in the right direction.
On Thu, Apr 1, 2010 at 1:50 PM, Thomas Watson <tjwatson@xxxxxxxxxx> wrote:
Note that Equinox does have the ability to
declare non-code dependencies in bundle manifests. See
Eclipse-GenericCapability and Eclipse-GenericRequire headers at:
This could be used by the DS implementation to declare a DS runtime capability and bundles defining DS components could declare a requirement on the DS runtime capability. But this mechanism only describes resolve time dependencies. It still would not solve Jeff's other concerns about the need to have DS active in order to truly work. Also note that p2 meta-data currently does not reflect the generic capabilities/requirements declared in a bundle manifest so even if we specified these today I don't think it would really help in ensuring a DS runtime is provisioned by p2. Perhaps we should consider adding that to p2?
Also note that the OSGi alliance is currently looking at providing a standard way for declaring generic capabilities and requirements for a future core specification. We should keep an eye on this space and feed any additional requirements we may have to OSGi in this area.
Jeff McAffer ---04/01/2010 11:46:32 AM---It should be up to the system integrator. Actually, there should be metadata (in p2) that expresses the need for various servic
It should be up to the system integrator. Actually, there should be metadata (in p2) that expresses the need for various services to be present to make the integrator's job easier but ultimately inclusion/activation/... are in the eye of the beholder. So we should not cod classpath (bundle or package) dependencies, rather we need more markup in p2 metadata to capture these non-classpath-related dependencies.
More detail: In this case you could declare a package dependency on the ds package but that will only get you the interfaces and not the implementation. The producer could similarly declare a bundle dependency on the Equinox ds bundle. This is short sighted as there are other DS implementations. Various p2 features could include the Equinox DS bundle. This is better but suffers from the same problem--that feature would not be usable with other DS implementations.
Note that the problem is a friend of the HTTP service, Help system and myriad of other situations where people need a service to be there but there is no clear declaration of that dependency.
Note also that simply having DS there is not enough. It needs to be started. This is a product/launch level concern (i.e., the DS bundle can/should not say that it should always be started).
So, unless the p2/ds problem is burning, it would be better to address the underlying issue than ad hoc addressing of the symptoms.
On 2010-04-01, at 12:21 PM, Susan Franklin McCourt wrote:
currently use ds in p2 to declare most of our services.
Yet we don't have any particular bundle that declares a dependency on ds.
I can justify this in some respects - theoretically there could be clients that consume the p2 bundles, declare their own services (using ds or dynamically) and thus don't care about getting the default service registrations. However this is not typical usage. Most people would expect to get the ds-declared services, and right now they only get cryptic errors or failed launches if they aren't using our features or product files and don't know to include ds.
For example, in a recent bug  , someone was getting a confusing error because we forgot to include ds in the .product file for an example. We fixed it by including ds in the product file.
My question is - is this the right fix?
It feels a little strange that a consumer that doesn't declare any services with ds still has to know that the bundles it is using declare their services this way.
Is it intentionally left up to the configurer of the system to ensure ds is included in the running target? Or should the bundles that declare services with ds be requiring the ds bundle?
equinox-dev mailing list
p2-dev mailing list