On Wed, Aug 24, 2011 at 7:55 PM, Christian Schneider <ch...@die-schneider.net> wrote: > Actually JDk and spring are two very good examples how to not do it :-) >
Yeah I guess using them as an example was "too easy" and it was late last evening. But often people would refer and look at giants such as the JDK and Spring. If you talk with the spring core team they always praise themselves as great API builders. And yeah they have succeeded IMHO to have a long line of releases with a stable API and making it very/fairly easy for end user to upgrade. For example many Spring 2.x apps would run fine in 3.x. Even older releases as well. But that said, the illusion with Spring is that its "all or nothing". So if you depend on one piece of Spring, you usually end up taking in X+ JARs. Just as the JMX example with Camel. I guess the Spring Expression Language would be another such example. And Spring is not as friendly for other framework builders, so there is not so many hooks we can leverage to influence spring. For example their property placeholder mechanism is a prime example. > I guess in the JDK no one cared as you will always have it. Btw. I guess > everyone agrees that the JDK is a mess architecturally. Btw. he JDK > extensions ship separate API jars like JAXB api. So they seem to have > learned. Yeah Oracle is looking into project jigsaw to try to refactor the JDK. That said the JDK has a remarkable backwards compatibility. Even so extreme that people have suggested to create a new Java, and dump the old cruft / @deprecated stuff etc. But that is another story that isn't very relevant in this thread. And of course Apache Camel should not be this extreme. Yes we have changed the API a bit between minor releases. Especially in the earlier releases of Camel 2.x where the API need to "settle down". But now in 9th release and going for 10th (Camel 2.9), the API ought to be stable. The changes you have done so far on trunk is IMHO good and fine. eg ServiceSupport into the support package. And have the old API marked as @deprecated. Even its not an absolute 100% backwards compatible change, as the impl is now using ServiceSupport from support package and not the old @deprecated one. So end users who oddly would type cast to ServiceSupport would have a type cast exception now. In fact one unit test in camel-ftp did that. But that is very uncommon for end users to do so I think the change is fine. > > In spring I suspect it is on purpose. They could provide API jars that make > you independent of their implementation. By combining API and impl they > force you into having a hard dependency on spring. > You had to remove the spring JMX annotations as we did not want to have > their impl. If they had cleanly separated their API from the impl we could > have kept the one API jar with the annotations and just implemented them > ourself when running outside of spring. > > So having the annotations in the management package is a very bad idea. A > subpackage would work on a pure simple package perspective but I think it > would be bad to have a top level package with implementations and a > subpackage with the API. > > We can move around the management stuff at the moment as my commit changed > it anyway. So before Camel 2.9 comes out we are free to move them. > > api.management of course only makes sense if we intend to put more stuff > there but I think it would be a good idea to do so. > Having a top level api package will also make it easier to create a pure API > jar for camel 3.0. I think it would be strange if the API jar would contain > > org.apache.camel > org.apache.camel.spi > org.apache.camel.management.annotation > > but not > org.apache.camel.management > > Btw management.annotation is not enough anyway as we have more management > interfaces that have to live in the API space. So management.api would be > better but I would prefer to have api at the top level so the user can > clearly see that everything api.* is part of the API. > > In any case we need to separate the management API from the management impl > classes. If we do not do it then we have no chance to avoid cycles. Besides > that how should we make it possible that the components only need to depend > on the API if we mix things. For example a component may want to use the > management annotations or another management interface but it should not > know the impl. > > Btw. the event classes should also be part of the API as they are necessary > to understand management events. As they live in a separate package already > the does not depend on the management impl I did not move them but they > would be better placed in api.management.events. > > Christian > > > > > Am 24.08.2011 19:12, schrieb Claus Ibsen: >> >> On Wed, Aug 24, 2011 at 6:17 PM, Christian Schneider >> <ch...@die-schneider.net> wrote: >>> >>> Hi Claus, >>> >>> we can do that but then we have to move the impl classes somewhere else. >>> We >>> may not mix impl and api in the same package. This is what leads to >>> cycles. >>> >> That is actually common. For example look at the JDK >> Map (API) and HashMap (Impl) are both in java.util package. >> >> However these annotations are not regular interfaces, that end users >> is supposed to implement. >> Or for example that we in the Apache Camel provides 2+ different >> implements of those annotations. >> >> As an end user I would feel natural these annotations are in the >> mangement package as they are part of the management >> (end user) API in Camel. >> >> >> The Spring framework put these annotations at >> >> http://static.springsource.org/spring/docs/3.0.x/javadoc-api/org/springframework/jmx/export/annotation/ManagedOperation.html >> >> We could also have a annotation subpackage >> (org.apache.camel.management.annotation) >> but we usually dont have that, eg there are no annotation package for >> @Consume, @Produce, @EndpointInject etc. >> >> Alternatively we could move them in the root package, but as you said >> there is already plenty of APIs in that package. >> >> Putting them in org.apache.camel.api seems a bit weird, as they would >> be the only pieces in there. >> And for Camel 2.x we should keep the API stable and not move around >> stuff all the time. >> >> >> >>> Christian >>> >>> >>> Am 24.08.2011 17:53, schrieb Claus Ibsen: >>>> >>>> On Wed, Aug 24, 2011 at 3:04 PM, Christian Schneider >>>> <ch...@die-schneider.net> wrote: >>>>> >>>>> So where do you propose to put them? >>>>> >>>>> 1. org.apache.camel >>>>> 2. org.apache.camel.api.management >>>>> >>>> I propose to put them here, where they where already >>>> 3. org.apache.camel.management >>>> >>>> These annotations are part of the management API in Camel and IMHO >>>> should be in that package. >>>> >>>> >>>> >>>>> I propose to go with 2 and create an api package with subpackages so we >>>>> can >>>>> structure org.apache.camel better. In the long run I would like to also >>>>> move >>>>> the whole camel api into an api package to make it clearer but that >>>>> will >>>>> probably create too much incompatibility. >>>>> >>>>> Christian >>>>> >>>>> >>>>> Am 24.08.2011 14:13, schrieb Claus Ibsen: >>>>>> >>>>>> On Tue, Aug 23, 2011 at 12:38 PM, Christian Schneider >>>>>> <ch...@die-schneider.net> wrote: >>>>>>> >>>>>>> I wonder what our scope for the org.apache.camel.spi package is vs >>>>>>> the >>>>>>> org.apache.camel (API) package. >>>>>>> >>>>>>> I know two valid definitions for API vs SPI: >>>>>>> >>>>>>> 1) API interfaces are called by the user to invoke functionality of >>>>>>> the >>>>>>> framework. So API interfaces are implemented by the framework. SPI >>>>>>> interfaces are implemented by the user to change functionality of the >>>>>>> framework or for callbacks >>>>>>> 2) SPI interfaces are for third party modules while API interfaces >>>>>>> are >>>>>>> for >>>>>>> users >>>>>>> >>>>>>> So the current case for me is the new JMX annotations. Are they SPI >>>>>>> interfaces or API interfaces? >>>>>>> >>>>>> They are API interfaces. Just like @Consumer, @Produce and any of the >>>>>> other API Camel annotations we have. >>>>>> Its just that these annotations is for management enabling your >>>>>> business logic / custom components or whatnot. >>>>>> >>>>>> >>>>>> >>>>>>> So what is your opinion about the specific and the general case. >>>>>>> >>>>>>> As a side question: The org.apache.camel package has grown quite >>>>>>> large. >>>>>>> I >>>>>>> think we should create specialized packages for it. As we are talking >>>>>>> about >>>>>>> the camel API org.apache.camel.api.* would be a good name in my >>>>>>> opinion. >>>>>>> So >>>>>>> the questions are: Should we create such specialized packages? Should >>>>>>> we >>>>>>> move API parts there? Should we only use the new packages for new >>>>>>> stuff? >>>>>>> >>>>>>> Christian >>>>>>> >>>>>>> >>>>>>> -- >>>>>>> -- >>>>>>> Christian Schneider >>>>>>> http://www.liquid-reality.de >>>>>>> >>>>>>> Open Source Architect >>>>>>> Talend Application Integration Division http://www.talend.com >>>>>>> >>>>>>> >>>>> -- >>>>> -- >>>>> Christian Schneider >>>>> http://www.liquid-reality.de >>>>> >>>>> Open Source Architect >>>>> Talend Application Integration Division http://www.talend.com >>>>> >>>>> >>>> >>> -- >>> Christian Schneider >>> http://www.liquid-reality.de >>> >>> Open Source Architect >>> http://www.talend.com >>> >>> >> >> > > -- > Christian Schneider > http://www.liquid-reality.de > > Open Source Architect > http://www.talend.com > > -- Claus Ibsen ----------------- FuseSource Email: cib...@fusesource.com Web: http://fusesource.com Twitter: davsclaus, fusenews Blog: http://davsclaus.blogspot.com/ Author of Camel in Action: http://www.manning.com/ibsen/
