Distributed OSGi example with Apache Karaf Cellar - Client bundle can't activate because can't find distributed service

Question

I am using Apache Karaf 4.1.1 and Karaf Cellar. I have written two bundles. The first bundle provides a service of type ITrackerManager. The second bundle has a component that references ITrackerManager. My end goal is to witness the component in the second bundle successfully get a reference to the ITrackerManager service in the first bundle which is running on a different node. This is all part of my exploration of distributed OSGi.

What is actually happening when I install that second bundle is that it gets installed but fails to activate due to missing the service reference. I must be conducting my test incorrectly. Any ideas on how I would go about demonstrating my end goal; component in bundle on Node B successfully uses service on Node A?

---

Here is how I have run my test so far.

Node A

karaf@root()> cluster:node-list
  | Id                | Alias | Host Name    | Port
--+-------------------+-------+--------------+-----
x | 159.4.251.58:5701 |       | 159.4.251.58 | 5701
  | 159.4.251.58:5702 |       | 159.4.251.58 | 5702

Node B

karaf@root()> cluster:node-list
  | Id                | Alias | Host Name    | Port
--+-------------------+-------+--------------+-----
  | 159.4.251.58:5701 |       | 159.4.251.58 | 5701
x | 159.4.251.58:5702 |       | 159.4.251.58 | 5702

So far so good. I am running two karaf instances on my computer. Both instances see each other. Now I want to install that first bundle onto Node A ONLY. To accomplish that, I install the bundle into the cluster, then specifically remove it from Node B.

Node A

karaf@root()> cluster:bundle-install -s default mvn:myCompany/dosgi-example-part1/1.0-SNAPSHOT

karaf@root()> cluster:bundle-list default
Bundles in cluster group default
ID | State    | Lvl | Located       | Blocked | Version        | Name
---+----------+-----+---------------+---------+----------------+--------------------------------------------------------------
 0 | Active   |     | cluster/local |         | 5.6.2          | System Bundle
...
67 | Active   |     | cluster/local       |         | 1.0.0.SNAPSHOT | Distributed OSGi Example Part 1

karaf@root()> cluster:service-list
Service Class             | Provider Node
--------------------------+------------------
myCompany.ITrackerManager | 159.4.251.58:5701
                          | 159.4.251.58:5702

Still looking good. My bundle is in the cluster, is local on Node A (and Node B at this point), and the service is recognized by the cluster and is available on both Node A and Node B. Now to remove the bundle from Node B.

Node B

karaf@root()> cluster:bundle-list default
Bundles in cluster group default
ID | State    | Lvl | Located       | Blocked | Version        | Name
---+----------+-----+---------------+---------+----------------+-------------------------------------------------------------
67 | Active   |     | cluster/local |         | 1.0.0.SNAPSHOT | Distributed OSGi Example Part 1

karaf@root()> bundle:list
START LEVEL 100 , List Threshold: 50
ID | State  | Lvl | Version        | Name
---+--------+-----+----------------+-----------------------------------------------
75 | Active |  80 | 1.0.0.SNAPSHOT | Distributed OSGi Example Part 1

karaf@root()> bundle:uninstall 75

karaf@root()> cluster:bundle-list default
Bundles in cluster group default
ID | State    | Lvl | Located       | Blocked | Version        | Name
---+----------+-----+---------------+---------+----------------+--------------------------------------------------------------
67 | Active   |     | cluster |         | 1.0.0.SNAPSHOT | Distributed OSGi Example Part 1

karaf@root()> cluster:service-list
Service Class             | Provider Node
--------------------------+------------------
myCompany.ITrackerManager | 159.4.251.58:5701

Excellent. The first bundle has been removed from Node B but still shows up as being in the cluster. Both nodes agree that my service is only available on Node A now (since the bundle was removed from Node B). Now I will load my second bundle on Node B only. This is where I run into problems. I don't load the second bundle using the cluster:bundle-install command because I don't want it ending up on Node A. So instead I install my second bundle using the normal bundle:install command. This results in an error about an unsatisfied reference.

Node B

karaf@root()> bundle:install -s mvn:otherCompany/dosgi-example-part2/1.0-SNAPSHOT
Bundle ID: 76
Error executing command: Error installing bundles:
        Unable to start bundle mvn:otherCompany/dosgi-example-part2/1.0-SNAPSHOT: org.osgi.framework.BundleException: Unable to resolve otherCompany.dosgi-example-part2 [76](R 76.0): missing requirement [otherCompany.dosgi-example-part2 [76](R 76.0)] osgi.wiring.package; (&(osgi.wiring.package=myCompany)(version>=1.0.0)(!(version>=2.0.0))) Unresolved requirements: [[otherCompany.dosgi-example-part2 [76](R 76.0)] osgi.wiring.package; (&(osgi.wiring.package=myCompany)(version>=1.0.0)(!(version>=2.0.0)))]

karaf@root()> bundle:list
START LEVEL 100 , List Threshold: 50
ID | State     | Lvl | Version        | Name
---+-----------+-----+----------------+-----------------------------------------------------------------------------------------------------
76 | Installed |  80 | 1.0.0.SNAPSHOT | Distributed OSGi Example Part 2

So there it is. I install the second bundle on NodeB only, expecting that it is able to successfully use the required service which resides on Node A only. Unfortunately that does not happen. Instead I get error message stating there are unresolved requirements. It seems to behave as if DOSGI is not available. If I install both bundles on the same node, the second bundle activates without any errors. Any insights you may have would be appreciated.

Answer 1

My problem was two-fold.

1. Stuff to be sent over DOSGI needs to be serializable. In my case, I was calling a method on a remote service that took an argument. That argument was a class type defined in a common API. That class type was not serializable. Once I made it serializable, it starting getting different errors. Which brings me to...

2. Normal name space rules apply. I will elaborate below.

My API defined two interfaces.

• ITracker
• ITrackerManager

That API bundle was installed into the cluster so it is available on all nodes. My Service bundle had a concrete implementation of ITrackerManager. When that bundle is installed locally on Node A, the cluster:service-list command correctly shows that Node A has a service of type ITrackerManager.

My Client bundle has a concrete implementation of ITracker that had a reference to ITrackerManager which was installed on Node B. The first thing the ITracker instance did in its activate method was call ITrackerManager.addTracker(this). What should have happened was that the instance of ITracker on Node B provided itself to the ITrackerManager running on Node A. Initially this failed because ITracker was not serializable. Once that was solved, I started seeing classNotFound exceptions on Node A.

Node A was trying to deserialize the ITracker instance locally. It was attempting to deserailize a concrete class (TheirTracker) which was not defined locally, it was only defined on Node B in the client bundle. This failed.

So the normal namespace rules apply. Even though the client bundle on Node B has a reference to a service running in a bundle Node A, the service bundle in Node A cannot create (i.e. deserialize) an instance of a class that only exists in the client bundle on Node B.

I switched up my interfaces so that ITrackerManager method does not take an ITracker arguement. Instead it takes a string. Invoking that method over DOSGi works fine.

While I understand why this problem exists, this undermines a core capability I was hoping to use with DOSGi. I want clients to be able to register with a central controller which will actively control them. This won't work because even though the clients implement the interface the central controller is looking for, the specific serialization fails at the central controller. The client concrete classes exist in a namespace unknown to the central controller, hence the client cannot successfully pass itself to the central controller.

This must be a way to achieve what I am looking for in DSOGi without making each of the multiple clients an exported DSOGi service. Any ideas?