I got quite confused by this discussion :-( and that makes it likely that more
people are.
Original question: when to track all References?
Rule 1: never. The only exception is when you’re doing something really special
and want to bypass the type system. This implies that you’re NOT bound to the
tracked type. Rule2: very rare, when tracking properties of the service are ok,
looking in its bundle might also work, but annotations would not work because
to work with the annotations you must be in the same class space for the
annotation packages. Basically if you track all references, everything is an
Object.
That said, I think there is a bit of confusion about class space compatibility.
The spec does not define compatibility, it defines incompatibility. Bundles are
incompatible when they would see more than one exporter for a package taking
the transitive closure of uses into account. So if either bundle has no import
of a package, then they are by definition not incompatible.
A: extendee
B: extender, will call A.register(foo.Foo), imports={foo}
C: tracker for Foo, exports={foo}
a) A.imports={}, so Foo is not incompatible for A.
b) A.imports={foo}, A, C, and B must be in the same class space for foo
As you should be able to see, if you just do the right thing (track only your
visible references) things work safely out of the box and you do not have to do
anything special.
So the remaining problematic scenario is when you want handle the case where
you have A.imports={foo} and A’s import is bound to a bundle D:
In this case, A & B are not in the same class space so B should NOT extend A.
This allows having multiple versions of A,B, and C. B could use A’s bundle to
load Foo, however, now C cannot see it (X2). It could use its own version but
then will get an error when registering it (X1). So you could now make C track
all services but then you get Foo objects that are not Foo objects to you, so
you need to handle them reflectively which utterly sucks and implies you no
longer have type safety. An easier solution is then to not import foo, then you
can see every object of the foo package.
Anyway, the moment you start to try to handle these cases your code will
balloon wildly out of proportions and become utterly unreadable. This rarely
worth it in my opinion because if you don’t want type safety, go to Javascript.
Kind regards,
Peter Kriens
> On 18 jun. 2015, at 21:46, [email protected] wrote:
>
> I haven't ever built a wicket extender, but I did once make a contraption
> which automagically generated a REST interface based on information in the
> target bundle. In this case B registers servlets which will be picked up
> by C, so far so good, but these servlets result in calls to classes which
> are defined in A - to which B is not normally wired. So B has to do some
> fancy stuff to load those packages (using A's class loader), but this does
> not imply that the servlets themselves should be registered on behalf of
> A: the servlets do not expose any of A's packages in the interfaces they
> register with C.
>
>> I agree with BJ.
>>
>> If the wicket extender were a wicket whiteboard, rather than an extender,
>> then I think we would expect it to track "wicket services" and then
>> register Servlet services using its own context as a result of finding
>> them. This is the same thing that lots of "adapter" bundles do.
>>
>> In the case of the wicket extender the only difference is how the wicket
>> object is obtained. Other than that it should behave in the same way as
>> the whiteboard version.
>>
>> Incidentally, it feels to me as if the wicket example would work better
>> with a whiteboard than an extender, as it would allow downstream service
>> dependencies for the wicket objects. I understand that it is only an
>> example though :)
>>
>> Tim
>>
>> Sent from my iPhone
>>
>>> On 18 Jun 2015, at 00:21, BJ Hargrave <[email protected]> wrote:
>>>
>>> I think it should be B, the wicket extender, since it is B that
>>> actually is wired to the servlet package and it is B which actually
>>> implements the Servlet. C does not implement the Servlet and does not
>>> import the servlet package. It just contributes implementation detail to
>>> the Servlet created by B.
>>>
>>> --
>>>
>>> BJ Hargrave
>>> Senior Technical Staff Member, IBM // office: +1 386 848 1781
>>> OSGi Fellow and CTO of the OSGi Alliance // mobile: +1 386 848 3788
>>> [email protected]
>>>
>>>
>>> ----- Original message -----
>>> From: Raymond Auge <[email protected]>
>>> Sent by: [email protected]
>>> To: OSGi Developer Mail List <[email protected]>
>>> Cc:
>>> Subject: Re: [osgi-dev] whiteboard pattern & extenders
>>> Date: Wed, Jun 17, 2015 5:54 PM
>>>
>>>
>>>
>>> On Wed, Jun 17, 2015 at 5:48 PM, BJ Hargrave <[email protected]>
>>> wrote:
>>> In this case B, the Wicket extender, must import the servlet package
>>> since it is making Servlet objects and registering them as Servlet
>>> services. It must use the same servlet package as A, the whiteboard
>>> impl, in order for A to understand the Servlet services.
>>>
>>> Ok, that's fine. Who's bundleContext should be used to register the
>>> service?
>>>
>>>
>>> --
>>>
>>> BJ Hargrave
>>> Senior Technical Staff Member, IBM // office: +1 386 848 1781
>>> OSGi Fellow and CTO of the OSGi Alliance // mobile: +1 386 848 3788
>>> [email protected]
>>>
>>>
>>> ----- Original message -----
>>> From: Raymond Auge <[email protected]>
>>> Sent by: [email protected]
>>> To: OSGi Developer Mail List <[email protected]>
>>> Cc:
>>> Subject: Re: [osgi-dev] whiteboard pattern & extenders
>>> Date: Wed, Jun 17, 2015 5:20 PM
>>>
>>> Haha, I think everyone is very close! But I will try very hard to be
>>> really really clear to the use case:
>>>
>>> Take Apache Wicket:
>>>
>>> https://wicket.apache.org/learn/examples/helloworld.html
>>>
>>> This frame work allows a developer to implement web applications without
>>> ever needing to touch the Servlet API. It's more like building native
>>> GUI building, except that it produces HTML.
>>>
>>> Most of the time you have to bundle all the framework jars (which
>>> contain the servlets).
>>>
>>> However, let's imagine that now the bundle with the wicket application
>>> only imports the wicket APIs (no framework jars, or servlet API).
>>>
>>> Now let's consider a Wicket extender. This bundle is the "wicket
>>> framework". It knows about how a wicket application should be
>>> bootstrapped. And it provides the concrete Servlet which exposes the
>>> application.
>>>
>>> Now the Wicket extender just wants to use the Http Whiteboard to
>>> register the wicket servlets.
>>>
>>> So, you have:
>>>
>>> A) the http whiteboard
>>> B) the Wicket Extender
>>> C) the Wicket application
>>>
>>> I can repeat this example many more times for many other web frameworks.
>>>
>>> - Ray
>>>
>>> On Wed, Jun 17, 2015 at 4:52 PM, Neil Bartlett <[email protected]>
>>> wrote:
>>> Felix,
>>>
>>>
>>>> On 17 Jun 2015, at 21:35, Felix Meschberger <[email protected]> wrote:
>>>>
>>>> Hi
>>>>
>>>>
>>>>> Am 17.06.2015 um 21:56 schrieb Neil Bartlett <[email protected]>:
>>>>>
>>>>> I think that B (the extender) must register the Servlet service using
>>>>> its own BundleContext, since it is the bundle that actually creates
>>>>> the Servlet objects.
>>>>
>>>> I donât think that works in general. And I actually think it is
>>>> wrong.
>>>
>>> No, I stand by it because your summary below doesnât match up with
>>> what Ray actually said. At least insofar as I have understood him
>>> correctly.
>>>
>>>>
>>>>
>>>> To repeat Rayâs example:
>>>>
>>>> (A) consumes a service, say javax.servlet.Servlet
>>>> (B) extends packages declaring something and registering services on
>>>> behalf of them
>>>> (C) declare something and provide the Servlets, hence implementations
>>>> of the javax.servlet.Servlet interface.
>>>
>>> Ray stated that the extended bundle C does NOT provide Servlets or know
>>> anything about Servlet API. It just creates these âwebby
>>> somethingsâ.
>>>
>>>>
>>>>
>>>> Now, C having the implementations implementing an interface *must* by
>>>> definition be wired to the service interface, otherwise the
>>>> implementations cannot be loaded by Câs class loader. And B must not
>>>> use its own (Bâs) class loader but must use Câs class loader to
>>>> load the implementations from C and use Câs bundle context to
>>>> register the service. B is only a messenger and Bâs bundle context
>>>> (and class loader) is never involved in this game. It cannot be
>>>> involved. Because it will, in general, never be able to load classes
>>>> from the extended bundle.
>>>
>>> Bâs classloader is involved because B makes the Servlet objects that
>>> wrap around whatever C provides.
>>>
>>> The way I understand this, C provides some kind of bean class, which may
>>> be a POJO. B instantiates that class (for this it would certainly have
>>> to use Câs classloader). It then creates a Servlet object that wraps
>>> around the POJO and forwards HTTP requests to it.
>>>
>>> Thus B registers the Servlet service using its own BundleContext. It
>>> imports javax.servlet, and the whiteboard will only pick up those
>>> Servlets if they comply with the same API version.
>>>
>>>
>>>>
>>>>
>>>> In any case for (A) to make normal use of the service provided by (C)
>>>> it must wire to the same service interface as (C) is wired to. Hence
>>>> (A) must not track all service references, hence using *false* on the
>>>> ServiceTracker to be able to make use of the Servlet service provided
>>>> by C (and instantiated and registered by B on behalf of C)
>>>>
>>>> This BTW actually *is* exactly the DS scenario, where the DS
>>>> implementation bundle would be B. The Http Service Whiteboard
>>>> implementation would be (A) and (C) is some bundle with a
>>>> Service-Components header.
>>>
>>> Well I disagree that itâs the same, for the reasons given above. So I
>>> guess Ray needs to come in here to clarify again.
>>>
>>>>
>>>>
>>>>
>>>>>
>>>>> Since the extended C bundle neither imports nor exports the Servlet
>>>>> API, *nobody* would be able to use its published Servlet services
>>>>> without doing trackAllServices=true. If you are required to turn on
>>>>> trackAllServices in your whiteboard bundle (A) then you are coupling
>>>>> that whiteboard to the implementation details of this service
>>>>> provider. All other potential consumers of your Servlet would have to
>>>>> do the same.
>>>>>
>>>>> The rule of thumb is that trackAllService is nearly always wrong
>>>>> unless you only need to inspect the metadata of a service without ever
>>>>> invoking it⦠for example if you are implementing a shell like Gogo.
>>>>
>>>> Agreed. There are some corner cases where trackAllServices makes sense,
>>>> but not in general (the Apache Aries JMX Whiteboard is another such use
>>>> case)
>>>>
>>>> Hope that helps more, than it confuses.
>>>>
>>>> Regards
>>>> Felix
>>>>
>>>>
>>>>>
>>>>>
>>>>> Regards,
>>>>> Neil
>>>>>
>>>>>> On 17 Jun 2015, at 20:47, Raymond Auge <[email protected]>
>>>>>> wrote:
>>>>>>
>>>>>>
>>>>>> On Wed, Jun 17, 2015 at 3:38 PM, BJ Hargrave <[email protected]>
>>>>>> wrote:
>>>>>> Well you were the one describing the scenario, I was trying to repeat
>>>>>> what I thought you were saying :-)
>>>>>>
>>>>>> So C is a bundle which does not have any implementation of Servlet
>>>>>> and does not import the servlet package and B will register a Servlet
>>>>>> service, using C's bundle context, with some object implementing
>>>>>> Servlet and B does not import the servlet package.
>>>>>>
>>>>>> How does B get an object implementing Servlet to register as the
>>>>>> service since it has no wiring to any package containing Servlet?
>>>>>>
>>>>>> I never said that B doesn't know about Servlet... In fact I said
>>>>>> exactly that B knows about making Servlets.
>>>>>>
>>>>>> It seems odd that neither B or C is wired to the servlet package, yet
>>>>>> they conspire to register a Servlet service.
>>>>>>
>>>>>> B should certainly be wired to the servlet package... and the same
>>>>>> one as the whiteboard.
>>>>>>
>>>>>> Let me try to clarify with a concrete example.
>>>>>>
>>>>>> There are many "webby" technologies in existence which remove the
>>>>>> need for a developer to have any knowledge of Servlet API. These
>>>>>> technologies use things like annotations or even simply pure
>>>>>> packaging conventions for describing their application.
>>>>>>
>>>>>> However, in the end, you need a servlet. Typically some framework
>>>>>> looks at the packaging convention and then reacts to that by creating
>>>>>> a Servlet which turns the convention into something concrete.
>>>>>>
>>>>>> In this scenario the original "bundle" doesn't know anything about
>>>>>> Servlet... BUT there is certainly a "concrete" servlet implementation
>>>>>> somewhere that knows about the convention.
>>>>>>
>>>>>> However, this concrete thing (the extender) wants to use the
>>>>>> whiteboard instead of handling all the HTTP stuff itself.
>>>>>>
>>>>>> the whiteboard knows nothing about this extender.
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>>
>>>>>> BJ Hargrave
>>>>>> Senior Technical Staff Member, IBM // office: +1 386 848 1781
>>>>>> OSGi Fellow and CTO of the OSGi Alliance // mobile: +1 386 848 3788
>>>>>> [email protected]
>>>>>>
>>>>>>
>>>>>> ----- Original message -----
>>>>>> From: Raymond Auge <[email protected]>
>>>>>> Sent by: [email protected]
>>>>>> To: OSGi Developer Mail List <[email protected]>
>>>>>> Cc:
>>>>>> Subject: Re: [osgi-dev] whiteboard pattern & extenders
>>>>>> Date: Wed, Jun 17, 2015 3:32 PM
>>>>>>
>>>>>> Actually Chris is correct in describing the scenario and BJ you are
>>>>>> not correct.
>>>>>>
>>>>>> C) is some bundle which has a header "ImCool: oh so cool!"
>>>>>> B) is an extender which makes servlets from the header "ImCool" IT
>>>>>> knows how to make a Servlet service.
>>>>>> A) is the whiteboard
>>>>>>
>>>>>>
>>>>>> This doesn't work because C) does not import Servlet.
>>>>>>
>>>>>> - Ray
>>>>>>
>>>>>> On Wed, Jun 17, 2015 at 3:24 PM, BJ Hargrave <[email protected]>
>>>>>> wrote:
>>>>>> OK.
>>>>>>
>>>>>> So A, the whiteboard impl, has ServiceTrackers and must care about
>>>>>> the specific package.
>>>>>>
>>>>>> B is the extends which registers the services. It has no
>>>>>> ServiceTrackers and does not care about the package since it does not
>>>>>> use the package itself.
>>>>>>
>>>>>> C also must care about the same package as A (so they are type
>>>>>> compatible).
>>>>>>
>>>>>> So there is not bundle which both is the extender and registers the
>>>>>> services and also has ServiceTrackers which must care about the
>>>>>> specific package. Therefore trackAllServices=true is not needed.
>>>>>>
>>>>>> --
>>>>>>
>>>>>> BJ Hargrave
>>>>>> Senior Technical Staff Member, IBM // office: +1 386 848 1781
>>>>>> OSGi Fellow and CTO of the OSGi Alliance // mobile: +1 386 848 3788
>>>>>> [email protected]
>>>>>>
>>>>>>
>>>>>> ----- Original message -----
>>>>>> From: Raymond Auge <[email protected]>
>>>>>> Sent by: [email protected]
>>>>>> To: OSGi Developer Mail List <[email protected]>
>>>>>> Cc:
>>>>>> Subject: Re: [osgi-dev] whiteboard pattern & extenders
>>>>>> Date: Wed, Jun 17, 2015 2:55 PM
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Wed, Jun 17, 2015 at 2:44 PM, BJ Hargrave <[email protected]>
>>>>>> wrote:
>>>>>> So this is like DS (an extender) registering Servlet services on
>>>>>> behalf of a bundle using DS. Then of course the extender bundle does
>>>>>> not care about the servlet package but also the extender bundle is
>>>>>> not using ServiceTrackers to track the Servlet services. That is done
>>>>>> by the Http Whiteboard impl bundle which does care about the servlet
>>>>>> package and its version.
>>>>>>
>>>>>> I'm sorry but you've lost me, and DS isn't an example of the scenario
>>>>>> because the DS bundle is itself tracker in this scenario.
>>>>>>
>>>>>> In the scenario I'm describing there are 3 bundles in play:
>>>>>>
>>>>>> A) the whiteboard bundle (has the trackers)
>>>>>> B) an extender which registers services that the whiteboard
>>>>>> C) a bundle which is being extended by B) but doesn't know anything
>>>>>> about A) or the API it's being extended with
>>>>>>
>>>>>> Sincerely,
>>>>>> - Ray
>>>>>>
>>>>>>
>>>>>> --
>>>>>>
>>>>>> BJ Hargrave
>>>>>> Senior Technical Staff Member, IBM office: +1 386 848 1781
>>>>>> OSGi Fellow and CTO of the OSGi Alliance mobile: +1 386 848 3788
>>>>>> [email protected]
>>>>>>
>>>>>>
>>>>>> ----- Original message -----
>>>>>> From: Raymond Auge <[email protected]>
>>>>>> Sent by: [email protected]
>>>>>> To: OSGi Developer Mail List <[email protected]>
>>>>>> Cc:
>>>>>> Subject: Re: [osgi-dev] whiteboard pattern & extenders
>>>>>> Date: Wed, Jun 17, 2015 2:23 PM
>>>>>>
>>>>>> But an extender who registers services to a whiteboard impl on behalf
>>>>>> of extendee will result in those services not being visible to the
>>>>>> whiteboard if the extendee does not import the packages used by the
>>>>>> services?
>>>>>>
>>>>>> On Wed, Jun 17, 2015 at 2:16 PM, BJ Hargrave <[email protected]>
>>>>>> wrote:
>>>>>> Well whiteboard and extenders are different.
>>>>>>
>>>>>> Whiteboard should not use true since it cares about the specific API
>>>>>> package version.
>>>>>>
>>>>>> Extenders should use BundleTrackers rather than ServiceTrackers since
>>>>>> they are not using whiteboard services.
>>>>>>
>>>>>> --
>>>>>>
>>>>>> BJ Hargrave
>>>>>> Senior Technical Staff Member, IBM office: +1 386 848 1781
>>>>>> OSGi Fellow and CTO of the OSGi Alliance mobile: +1 386 848 3788
>>>>>> [email protected]
>>>>>>
>>>>>>
>>>>>> ----- Original message -----
>>>>>> From: Raymond Auge <[email protected]>
>>>>>> Sent by: [email protected]
>>>>>> To: OSGi Developer Mail List <[email protected]>
>>>>>> Cc:
>>>>>> Subject: [osgi-dev] whiteboard pattern & extenders
>>>>>> Date: Wed, Jun 17, 2015 2:12 PM
>>>>>>
>>>>>> When implementing a whiteboard pattern should we always open trackers
>>>>>> using the trackAllServices = true ? via:
>>>>>>
>>>>>> ServiceTracker.open(true);
>>>>>>
>>>>>> It would seem that this is the only way that we can support extenders
>>>>>> where the extendee has no knowledge of the APIs in question, correct?
>>>>>>
>>>>>> --
>>>>>> Raymond Augé (@rotty3000)
>>>>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>>>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Raymond Augé (@rotty3000)
>>>>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>>>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Raymond Augé (@rotty3000)
>>>>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>>>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Raymond Augé (@rotty3000)
>>>>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>>>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>>>
>>>>>>
>>>>>>
>>>>>> --
>>>>>> Raymond Augé (@rotty3000)
>>>>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>>>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>>>>> _______________________________________________
>>>>>> OSGi Developer Mail List
>>>>>> [email protected]
>>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>>> _______________________________________________
>>>>> OSGi Developer Mail List
>>>>> [email protected]
>>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>> _______________________________________________
>>>> OSGi Developer Mail List
>>>> [email protected]
>>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>
>>> _______________________________________________
>>> OSGi Developer Mail List
>>> [email protected]
>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>
>>>
>>>
>>> --
>>> Raymond Augé (@rotty3000)
>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>> _______________________________________________
>>> OSGi Developer Mail List
>>> [email protected]
>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>
>>>
>>> _______________________________________________
>>> OSGi Developer Mail List
>>> [email protected]
>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>
>>>
>>>
>>> --
>>> Raymond Augé (@rotty3000)
>>> Senior Software Architect Liferay, Inc. (@Liferay)
>>> Board Member & EEG Co-Chair, OSGi Alliance (@OSGiAlliance)
>>> _______________________________________________
>>> OSGi Developer Mail List
>>> [email protected]
>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>>>
>>> _______________________________________________
>>> OSGi Developer Mail List
>>> [email protected]
>>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>> _______________________________________________
>> OSGi Developer Mail List
>> [email protected]
>> https://mail.osgi.org/mailman/listinfo/osgi-dev
>
>
> _______________________________________________
> OSGi Developer Mail List
> [email protected]
> https://mail.osgi.org/mailman/listinfo/osgi-dev
_______________________________________________
OSGi Developer Mail List
[email protected]
https://mail.osgi.org/mailman/listinfo/osgi-dev
