Re: [equinox-dev] Question about DI/ DS and Application model
BJ Hargrave wrote on 09/17/2009 10:41:58 AM: because services from lazy activated bundles are available prior to being started Technically that is not true. Services from lazy activated bundles are available prior to being *activated* but the bundle must have been started. That is, not in RESOLVED state; STARTING or ACTIVE. This is just me mixing up the bundle lifecycle terminology. By started I meant active (i.e., the state that comes after starting). I guess there is a different meaning to the term started in OSGi that I wasn't aware of. Thanks for clarifying. Since the bundle providing the extension is STARTED at this point, and all other lazy activated bundles are STARTING This sounds like a start ordering issue. Since extensions are active in the RESOLVED state, the system will need to be configured such that all bundles which will use extension to access services are started before they will ever attempt to access the service. Extensions are present in the extension registry in the RESOLVED state, but an executable extension will only start to use services once it has been instantiated. I'm assuming that bundles are lazy-activated here and at the time the extension class is instantiated the bundle is active. In any case, there may indeed be lifecycle issues here but I think I need to play around some more with combining extensions and DS to see if there is a problem. John___ equinox-dev mailing list equinox-dev@eclipse.org https://dev.eclipse.org/mailman/listinfo/equinox-dev
Re: [equinox-dev] Question about DI/ DS and Application model
Pascal, I think there are deeper problems here than just the lifecycle. Neither DS nor the Eclipse extension registry were particularly designed to work together. In DS objects are created for only two purposes: (a) To be published to the world as a service, and/or (b) to have an independent lifecycle delimited by activate/deactivate method, giving the object the opportunity to, say, open a socket, run a polling thread, etc. If an object is neither published nor active, then there is no point in DS creating it! The extension registry is not so different. Let's disregard pure metadata extensions that do not have some kind of class attribute. An extension is a little factory that is published to the world much like services are, albeit with some additional notions of scope provided by extension point IDs. Extensions do not have an independent lifecycle, so their only purpose is to be published. These worlds are separate. Imagine for a moment trying to do the opposite of what you ask: invoke an extension-contributed class from a DS-instantiated component. The component cannot be injected with the extension, so it would have to go to the extension registry explicitly and run a query just like everybody else. Now flip it back to your problem, the extension registry always instantiates objects directly or via an IExecutableExtensionFactory (IEEF); we cannot tell it to pick a service instead. So you have to go to the service registry explicitly and run a query. Actually there are tricks we can perform with IEEF that will allow us to grab services, by specifying an OSGi service filter string in the configuration data for the IEEF. I have experimented with this approach and actually gave some demos at EclipseCon 07. But there is a further impedence mismatch: extensions are factories and consumers of extensions expect to get a new object each time, whereas services are reused. Try reusing an IPreferencePage object... you'll get SWT Widget Disposed errors the second time you open the prefs dialog. An alternative approach that might work for you is to use the new Blueprint Container. In Blueprint, not every bean needs to be published as a service, but the container itself for a bundle is published as a service and allows access to the beans within. Martin Lippert wrote a reusable IEEF that worked exactly this way with Spring-DM before Blueprint existed; I'm sure it can be updated to the new API. Furthermore Blueprint solves the problem of reuse of objects: any bean can be declared with prototype scope, turning it into a mini-factory where a new instance is created on each request. Neil On Thu, Sep 17, 2009 at 7:46 PM, Pascal Rapicault pascal_rapica...@ca.ibm.com wrote: I agree that for this particular example where only one service is acquired the usage of DS is probably overkill. However if several services had been needed, I think it would have probably been better to use DS over writing all the service handling manually. As for the relevance of the example, I would argue that it is very relevant as this is probably the simplest thing that one can do: write a little piece of code that uses services to perform a task and then maybe make it available as a application (or an action, or...). It does not need nor want to publish a service. At this point the two root causes seem to be: - The only way for DS to provide the objects resulting from the injection is through services and this is cumbersome and ends up polluting the service space. - Bridging the service and extension world so that created extensions can get things injected to them. I looked at the code in the solution3 package and if I understand it correctly, it seems to be an implementation of 1.2 since the RepositoryDumper is registered by DS and then acquired by the application itself. PaScaL [image: Inactive hide details for John Arthorne---09/16/2009 09:58:08 PM---Eventually someone has to decide which implementation of IMe]John Arthorne---09/16/2009 09:58:08 PM---Eventually someone has to decide which implementation of IMetadataRepositoryManager is going to be used. I think in the case of From: John Arthorne/Ottawa/i...@ibmca To: Equinox development mailing list equinox-dev@eclipse.org Date: 09/16/2009 09:58 PM Subject: Re: [equinox-dev] Question about DI/ DS and Application model -- Eventually someone has to decide which implementation of IMetadataRepositoryManager is going to be used. I think in the case of an application it is quite reasonable for the application to make this decision directly (by looking up the service, perhaps with some filter that helps to select the manager to use). By moving the lookup of IMetadataRepositoryManager into a DS component it just hides the fact that it is a simple service lookup and doesn't seem to offer any advantage. I think because both the service declaration, the implementation, and the client are all in the same bundle
Re: [equinox-dev] Question about DI/ DS and Application model
Eventually someone has to decide which implementation of IMetadataRepositoryManager is going to be used. I think in the case of an application it is quite reasonable for the application to make this decision directly (by looking up the service, perhaps with some filter that helps to select the manager to use). By moving the lookup of IMetadataRepositoryManager into a DS component it just hides the fact that it is a simple service lookup and doesn't seem to offer any advantage. I think because both the service declaration, the implementation, and the client are all in the same bundle it's not a particularly interesting case. However I could imagine in more complex cases something like your solution 3 would be interesting. An executable extension factory could allow the services required by an executable extension to be injected into it rather than having the extension reach out. You'll see another package solution3 in the bundle where I was playing around with another approach. I'm not sure it's any better than your solution 1 but you can take a look. John Pascal Rapicault/Ottawa/i...@ibmca Sent by: equinox-dev-boun...@eclipse.org 09/16/2009 04:00 PM Please respond to Equinox development mailing list equinox-dev@eclipse.org To Equinox development mailing list equinox-dev@eclipse.org cc Subject [equinox-dev] Question about DI/ DS and Application model Today I have done some more DI exploration using DS to see how it fits with the constructs we have in eclipse and I'm struggling to integrate in a nice way with the application model (I mean without using static) and I'm looking to know how others are doing this? The one line summary of my experiment is: I have a class that does some work (named RepositoryDumper), it needs a service (RepoMgr). I want now to create an eclipse application that invokes the RepositoryDumper and I would like to not have to acquire the RepoMgr service manually. Here is what I have been exploring with: Solution 1: I have an application declared in the plugin.xml. I have created a DS component that instantiates RepositoryDumper. However the question now is how does the application (remember that an eclipse application extension needs to provides n class) can get a hold of the RepositoryDumper instance that got created by DS: - 1.1: Ugly - Store the instance RepositoryDumper in the Activator of the plug-in - 1.2: Get the RepositoryDumper be registered as a Service and have the application get this service. I don't like this because now RepositoryDumper is visible to everybody just so I can get access to it Solution 2: This solution assumes that the declarative approach to the eclipse application model is the hindrance and works around it by registering an ApplicationDescriptor (org.osgi.service.application). To do so I create a DS component that instantiates the RepositoryDumper and also register an ApplicationDescriptor as a service. This has the nice attribute that everything gets injected and that the application is only available to run if all the necessary pieces are available. However it requires a lot of code since one has to implement ApplicationDescriptor and ApplicationHandle, and I don't think this application would even be launchable using the -application argument. Solution 3: This solution is an hybrid between 1 and 2 using the IExecutableExtensionFactory. There is a DS component that creates the RepositoryDumper and register a service, let's call it X. Then let's make the class specified in application extension (in the plugin.xml) implements IExecutableExtensionFactory and have it get the service X. This solution allows to have the application construction be completely done by injection however given that the application is contributed through extension registry it still is visible even though not ready to run. How are others doing this? Is this a real problem or is it just me? Should I just not worry about that and use static fields? Btw, the code is available /cvsroot/rt org.eclipse.equinox/incubator/p2/bundles/org.eclipse.equinox.p2.diagnostic Only solution 1 and 2 are available. Thx for your attention and feedback PaScaL___ equinox-dev mailing list equinox-dev@eclipse.org https://dev.eclipse.org/mailman/listinfo/equinox-dev ___ equinox-dev mailing list equinox-dev@eclipse.org https://dev.eclipse.org/mailman/listinfo/equinox-dev
Re: [equinox-dev] Question about DI/ DS and Application model
Isn't there a big problem with the life cycle mismatch between services and extensions? Services require a bundle to be started. Extensions require a bundle to be resolved. -- BJ Hargrave Senior Technical Staff Member, IBM OSGi Fellow and CTO of the OSGi Alliance hargr...@us.ibm.com office: +1 386 848 1781 mobile: +1 386 848 3788 From: John Arthorne john_artho...@ca.ibm.com To: Equinox development mailing list equinox-dev@eclipse.org Date: 2009/09/16 21:58 Subject: Re: [equinox-dev] Question about DI/ DS and Application model Sent by: equinox-dev-boun...@eclipse.org Eventually someone has to decide which implementation of IMetadataRepositoryManager is going to be used. I think in the case of an application it is quite reasonable for the application to make this decision directly (by looking up the service, perhaps with some filter that helps to select the manager to use). By moving the lookup of IMetadataRepositoryManager into a DS component it just hides the fact that it is a simple service lookup and doesn't seem to offer any advantage. I think because both the service declaration, the implementation, and the client are all in the same bundle it's not a particularly interesting case. However I could imagine in more complex cases something like your solution 3 would be interesting. An executable extension factory could allow the services required by an executable extension to be injected into it rather than having the extension reach out. You'll see another package solution3 in the bundle where I was playing around with another approach. I'm not sure it's any better than your solution 1 but you can take a look. John Pascal Rapicault/Ottawa/i...@ibmca Sent by: equinox-dev-boun...@eclipse.org 09/16/2009 04:00 PM Please respond to Equinox development mailing list equinox-dev@eclipse.org To Equinox development mailing list equinox-dev@eclipse.org cc Subject [equinox-dev] Question about DI/ DS and Application model Today I have done some more DI exploration using DS to see how it fits with the constructs we have in eclipse and I'm struggling to integrate in a nice way with the application model (I mean without using static) and I'm looking to know how others are doing this? The one line summary of my experiment is: I have a class that does some work (named RepositoryDumper), it needs a service (RepoMgr). I want now to create an eclipse application that invokes the RepositoryDumper and I would like to not have to acquire the RepoMgr service manually. Here is what I have been exploring with: Solution 1: I have an application declared in the plugin.xml. I have created a DS component that instantiates RepositoryDumper. However the question now is how does the application (remember that an eclipse application extension needs to provides n class) can get a hold of the RepositoryDumper instance that got created by DS: - 1.1: Ugly - Store the instance RepositoryDumper in the Activator of the plug-in - 1.2: Get the RepositoryDumper be registered as a Service and have the application get this service. I don't like this because now RepositoryDumper is visible to everybody just so I can get access to it Solution 2: This solution assumes that the declarative approach to the eclipse application model is the hindrance and works around it by registering an ApplicationDescriptor (org.osgi.service.application). To do so I create a DS component that instantiates the RepositoryDumper and also register an ApplicationDescriptor as a service. This has the nice attribute that everything gets injected and that the application is only available to run if all the necessary pieces are available. However it requires a lot of code since one has to implement ApplicationDescriptor and ApplicationHandle, and I don't think this application would even be launchable using the -application argument. Solution 3: This solution is an hybrid between 1 and 2 using the IExecutableExtensionFactory. There is a DS component that creates the RepositoryDumper and register a service, let's call it X. Then let's make the class specified in application extension (in the plugin.xml) implements IExecutableExtensionFactory and have it get the service X. This solution allows to have the application construction be completely done by injection however given that the application is contributed through extension registry it still is visible even though not ready to run. How are others doing this? Is this a real problem or is it just me? Should I just not worry about that and use static fields? Btw, the code is available /cvsroot/rt org.eclipse.equinox/incubator/p2/bundles/org.eclipse.equinox.p2.diagnostic Only solution 1 and 2 are available. Thx for your attention and feedback PaScaL___ equinox-dev mailing list equinox-dev@eclipse.org https://dev.eclipse.org/mailman/listinfo/equinox-dev