Only activate OSGi component when an optional dependency is available - osgi

I have an OSGi bundle that declares an optional dependency on the org.osgi.service.subsystem package. It listens for the addition of a new Subsystem instance, using DS. However, this DS component activates even if the dependency is unavailable.
Is there a way to only activate this DS component if the optional dependency is available?
Basically, this DS component performs processing that is only relevant if there are actually subsystems present in the runtime. If there are none, and the class is not available in the runtime, this component should remain dormant.

As far as I know this is not possible. There is an even bigger problem that you bundle will not automatically bind to the optional package if it becomes available after you bundle resolved. You will have to refresh the bundle explicitly.
So the better solution in this case would be to move the component into a separate bundle that has a mandatory dependency on the subsystem package. It could then be installed together with the subsystem support.

Related

java.util.ServiceConfigurationError Provider not a subtype while using OSGi bundle

I'm creating a Liferay 7.1 OSGi bundle, which has some external dependencies in it. In consideration of time, we opted to embed the external JAR in our OSGi Bundle. I've managed to create a bnd file, which includes all of the ElasticSearch dependencies, and put them on the bundle classpath. I've used the source-code from github (https://github.com/liferay/liferay-portal/blob/master/modules/apps/portal-search-elasticsearch6/portal-search-elasticsearch6-impl/build.gradle) and the bnd.bnd file, to check what's imported.
When activating the bundle, an exception is thrown:
The activate method has thrown an exception
java.util.ServiceConfigurationError: org.elasticsearch.common.xcontent.XContentBuilderExtension: Provider org.elasticsearch.common.xcontent.XContentElasticsearchExtension not a subtype
at java.util.ServiceLoader.fail(ServiceLoader.java:239)
at java.util.ServiceLoader.access$300(ServiceLoader.java:185)
at java.util.ServiceLoader$LazyIterator.nextService(ServiceLoader.java:376)
at java.util.ServiceLoader$LazyIterator.next(ServiceLoader.java:404)
at java.util.ServiceLoader$1.next(ServiceLoader.java:480)
at org.elasticsearch.common.xcontent.XContentBuilder.<clinit>(XContentBuilder.java:118)
at org.elasticsearch.common.settings.Setting.arrayToParsableString(Setting.java:1257)
The XContentBuilderExtension is from the elasticsearch-x-content-6.5.0.jar,
the XContentElasticsearchExtension class, is included in the elasticsearch-6.5.0.jar. Both are Included Resources, and have been put on the classpath.
The Activate-method initializes a TransportClient in my other jar, hence it happens on activation ;).
Edit:
I've noticed that this error does NOT occur when installing this the first time, or when the portal restarts. So it only occurs when I uninstall and reinstall the bundle. (This is functionality I really prefer to have!). Maybe a stupid thought.. But could it be that there is some 'hanging thread'? That the bundle is not correctly installed, or that the TransportClient still is alive? I'm checking this out. Any hints are welcome!
Edit 2:
I'm fearing this is an incompatibility between SPI and OSGi? I've checked: The High Level Rest Client has the same issue. (But then with another Extension). I'm going to try the Low-Level Rest Client. This should work, as there are minimal dependencies, I'm guessing. I'm still very curious on why the incompatibility is there. I'm certainly no expert on OSGi, neither on SPI. (Time to learn new stuff!)
Seems like a case where OSGi uses your bundle to solve a dependency from another bundle, probably one that used your bundle to solve a package when the system started.
Looking at the symptoms: it does not occur when booting or restarts. Also it is not a subtype.
When OSGi uses that bundle to solve a dependency, it will keep a copy around, even when you remove it. When the bundle comes back a package that was previously used by another bundle may still be around and you can have the situation where a class used has two version of itself, from different classloaders, meaning they are not the same class and therefore, not a subtype.
Expose only the necessary to minimize the effects of this. Import only if needs importing. If you are using Liferay Gradle configuration to include the bundle inside, stop - it's a terrible way to include as it exposes a lot. If using the bnd file to include a resource and create an entry for the adicional classpath location, do not expose if not necessary. If you have several bundles using one as dependency, make sure about the version they use and if the exchange objects from the problematic class, if they do, than extra care is required.
PS: you can include attributes when exporting and/or importing in order to be more specific and avoid using packages from the wrong origin.
You can have 2 elastic search connections inside one Java app and Liferay is by default not exposing the connection that it holds.
A way around it is to rebuild the Liferay ES connector. It's not a big deal because you don't need to change the code only the OSGi descriptor to expose more services.
I did it in one POC project and worked fine. The tricky thing is to rebuild the Liferay jar but that was explained by Pettry by his google like search blog posts. https://community.liferay.com/blogs/-/blogs/creating-a-google-like-search (it is a series but it's kind of hard to navigate in the new Liferay blogs but Google will probably help) Either way it is all nicely documented here https://github.com/peerkar/liferay-gsearch
the only thing then what needs to be done is to add org.elasticsearch.* in the bnd.bnd file in the export section. You will then be able to work with the native elastic API.

Use restricted class in an OSGi Bundle

For publishing with SSL using the Endpoint I need to access classes under the packages com.sun.net.httpserver.*
Using the Eclipse IDE I found a way to use this classes. But exporting the bundles and running them in another equinox OSGi Installation I can't start the bundle throwing the following error:
java.lang.NoClassDefFoundError: com/sun/net/httpserver/HttpsConfigurator
Anyone an Idea how to solve this issue?
Thanks!
The package you're referring to is part of the JDK. You need to expose it, to make it available in OSGi and you have two options:
The first, and in most cases preferred option, is to expose this package through the system bundle. The OSGi framework has a property that you can set to do this:
org.osgi.framework.system.packages.extra=...
As its value, you provide it with a comma separated list of packages that you want to expose, on top of the ones that are already exposed by the framework. In your case, at least com.sun.net.httpserver, but there might be more packages that you need. In this case, also make sure that the bundle that uses this package imports this package.
The second option is to use a mechanism used boot delegation. It should only be used as a last resort, as it breaks modularity and if it's not used carefully it might lead to other problems. Again, this is a property that you need to set:
org.osgi.framework.bootdelegation=*
Here, you can provide a comma separated list of packages that should be loaded by the boot class loader. Wildcards are supported (as seen in the example above) but you are encouraged to be as specific as possible, so in your case for example use com.sun.* as the value.

When a bundle is programmatically started, should dependencies also be started?

Here's the scenario:
I have a bundle 'BundleA' installed and started in an OSGi container. A new version of BundleA is available. BundleA is provisioned using Felix Bundle Repository. The new version of BundleA declares a new package level requirement on a package from 'BundleB'.
Before I update BundleA (using OBR's deploy()) I stop BundleA because I want all threads to stop running and the deactivator provides this ability.
When I perform a deploy() on BundleA, BundleB is also installed, as expected.
I then programmatically start() BundleA again, and BundleA starts. But BundleB is 'resolved', not 'active'. I can manually start BundleB and it works as expected.
Is this expected, related to a way I am programmatically calling OSGi API, or did something go wrong?
That's the expected default behavior.
You can enable automatic activation of bundles as soon as any class from them is loaded. In order to do so, you need to set the Bundle-ActivationPolicy: lazy header.
In Eclipse that's the checkbox "Activate this plug-in when one of its classes is loaded" in the Manifest editor on the Overview page.
The OSGi framework does not automatically start bundles. (It will restart previously started bundles on framework launch however.) The framework does not know anything about start dependencies between bundles and bundles should not require a certain start order. As Gunnar mentioned, you can use activation policy to trigger lazy activation but that will not do anything if you have not called start on the bundle with the lazy activation policy.

What is the natural start order for package-dependent OSGI bundles (under Karaf)?

I have a problem on 2.2.8 version of Karaf (and most probably on the earlier versions too).
I'm going to use Karaf to host the system with dynamically deployed bundles. Bundles are deployed by users and i cannot know beforehand which are they.
I expect order of the BundleActivator.start() to exactly correspond to package dependencies between bundles (dependencies of import/export packages) and planning to expect that it will be safe to assume that bundle0 will be completely initialized before bundle1 is going to be started. But it is not so - it seems that BundleActivator.start() is invoked in a "random" order and disregards package dependencies between bundles.
Sample use-case, I have 3 libs
test-lib0 - defines testlib0.ITestRoot, exports testlib0 package
test-lib1 - defines testlib1.TestRoot implements ITestRoot, exports testlib1 package
test-lib2 - uses both libs, ITestRoot and TestRoot
When Karaf is started, i see following sample output in console
karaf#root> TestLib1Activator.start()
TestLib2Activator.start()
ITestRoot: interface com.testorg.testlib0.ITestRoot - 16634462
TestRoot: class com.testorg.testlib1.TestRoot - 21576551
TestLib0Activator.start()
but i expect it should be always in this order
TestLib0Activator.start()
TestLib1Activator.start()
TestLib2Activator.start()
ITestRoot: interface com.testorg.testlib0.ITestRoot - 16634462
TestRoot: class com.testorg.testlib1.TestRoot - 21576551
I'm attaching sample project for tests. Test case: after "mvn install" just move jars from ./deploy folder to the same folder of Karaf, trace messages should appear in console.
(Note: it may work correctly from the first attempt, try one more time then :))
Sample test project
http://karaf.922171.n3.nabble.com/file/n4025256/KarafTest.zip
Note: this is cross-post from http://karaf.922171.n3.nabble.com/What-is-the-natural-start-order-for-dependent-bundle-td4025256.html
In OSGi the bundle lifecycle is installed → resolved → starting → started.
Import-Package and Export-Package only influence when the bundle goes from installed to resolved. So the framework makes sure all bundles you import packages from are resolved before your bundle but then your bundle only goes to the resolved state. Then in a second step the activators are called. So you can not assume the activators are called in the same order. If you need some initializations before your testlib2 can work then you should use OSGi services.
So If I understood your case correctly then you testlib0 defines an interface, testlib1 implements it and testlib2 wants to use the implementation. So the best way to achieve this is to publish the impl as an OSGi service in testlib1 and reference this service in testlib3. You can then use the service with a ServiceTracker or with e.g. blueprint. I have a small example that shows this: http://www.liquid-reality.de/x/DIBZ . So if you do your case like in my example blueprint makes sure that the context of testlib2 only gets started when the service is there. It will even stop testlib2 when the service goes away.

Is separate OSGI bundles for api and implementation common practice?

I have a class with dependencies which I want to hot deploy without restarting the dependencies. The class has an interface but there's only one concrete implementation.
Initially I created a single bundle with exported the interface and registered the implementation using activator and implementation classes which were not exported. However, if I update the bundle, bundles which use the registered service get restarted after the update when PackageAdmin#refreshPackages is called (this is automatic when using fileinstall).
I have fixed this by creating a separate api bundle.
Is this the best way to achieve this?
Would you ever have a bundle which exports its api and includes the implementation in the same bundle. As far as I can see any give bundle would either export all its classes or no classes. What am I missing?
It is definitely a best practice to separate API bundles from their implementations in OSGi. If you do this, then any bundle that uses the API only needs to import classes from the API bundle, which can allow you to change implementations at runtime without restarting your dependent bundles.
Ideally your implementation bundle would implement the interface and export implementation as a service on the API provided interface. This allows the client bundles to utilize the service without referencing the implementation bundle.
In Apache Sling we do both: major APIs are in their own bundles, but for smaller things like extensions or optional components we often provide the default implementation in the same bundle as the API.
In the latter case, you can still allow for those default services to be replaceable, for example using service ranking values when you want to override them.
A bundle does not have to export all its classes, our bundles which include default services export just the API packages (and the default implementations are in different packages, obviously).
Unless there is a hard requirement to be able to replace implementation at runtime, without restarting client bundles, I would personally advocate keeping the explicit dependency link between API and implementation (either by including impl classes in the API bundle, or by having the API bundle import implementation packages from the impl. bundle).
The pb with the patterns suggested above is that they break the dependency chain. The benefits of dependency management go far beyond simple API compatibility, they also include ensuring predictable, consistent runtime behavior, as well as compatibility with the deployment ecosystem, and all of those require managing the implementation dependencies.

Resources