Skip to main content
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] [List Home]
Re: [ecf-dev] remote services, load balancing 
Hi Greg,

Greg House wrote:

Hello,
I am moving the discussion (http://www.eclipse.org/forums/index.php?t=rview&goto=528121&th=166590#msg_528121 <http://www.eclipse.org/forums/index.php?t=rview&goto=528121&th=166590#msg_528121>) to the mailing list as requested by Scott.
Sorry if this e-mail is too long. I will explain the use-case I am trying to implement.
Consider two or more OSGi platforms. Each platform exports (remote) services that the other platforms can use. At any time a new platform can appear (or disappear) and the other platforms should be able to see and use the new services without any additional 
configuration or restart.
Because some of the services are expensive and time-consuming operations I want to be able to deploy them on several machines and the requests to these services to be load-balanced. In addition, I want to be transparent for a consumer if a local or a remote service is used. I don't want in my code to import any OSGi packages.
For the platform and remote services implementations I am using Equinox and ECF.
Now I know that if we remove the load-balancing requirement all this can be implemented with (remote) declarative services. Moreover, the distribution and discovery providers can be independently selected.
However, the OSGI specification does not say anything about load-balancing, so one has to use some “nonstandard” approach and stick to specific distribution and discovery providers (and external components) – is my understanding correct?
Yes. 

The ECF has a solution to load-balance remote-services with JMS/ActiveMQ.
For my use-cases it is OK to use ECF ActiveMQ provider and maintain separate ActiveMQ installations. However, is it possible the usage of services to be transparent for the consumers? The service providers define the properties (local, remote, load-balanced) 
of their implementation and the consumers just use the services?
Yes, this is correct. The load-balanced service usage can be transparent for the consumers.
I tried to convert the org.eclipse.ecf.examples.loadbalancing.* examples to declarative services and to remove all ECF references from the code 
but with no success.
What went wrong? I don't immediately see a reason why this wouldn't work...for the loadbalancing consumer/client.
The conversion of org.eclipse.ecf.examples.loadbalancing.server to declarative services is straightforward.
About org.eclipse.ecf.examples.loadbalancing.servicehost I created the following OSFI-INF/component.xml: 

<?xml version="1.0" encoding="UTF-8"?>

   <implementation class="null" />
   <property name="service.exported.interfaces" type="String" value="*"/>
<property name="service.exported.configs" type="String" value="ecf.jms.activemq.tcp.manager.lb.svchost"/> 
   <property name="org.eclipse.ecf.containerFactoryArgs">
       tcp://192.168.1.1:61616/exampleQueue
       tcp://192.168.1.1:61616/exampleTopic
   </property>
<property name="ecf.rsvc.proxy" type="String" value="true"/> <service> <provide interface="org.eclipse.ecf.examples.loadbalancing.IDataProcessor"/> 
   </service>
<reference cardinality="1..1" interface="org.eclipse.ecf.core.IContainerFactory" name="IContainerFactory" policy="static"/> 
</scr:component>
Should I use the service.exported.interfaces property? Is <implementation class="null" /> correct?
I would have to check the ds spec, but no, I don't think that 'null' is allowable for the implementation class.
The main problem I have is with org.eclipse.ecf.examples.loadbalancing.consumer. 
I created the following OSGI-INF/component.xml file:

<?xml version="1.0" encoding="UTF-8"?>
<implementation class="org.eclipse.ecf.internal.examples.loadbalancing.consumer.DataProcessorConsumerApplicationDS"/> <reference bind="doTests" cardinality="0..n" interface="org.eclipse.ecf.examples.loadbalancing.IDataProcessor" name="IDataProcessor" policy="dynamic"/> 
</scr:component>
Is this somehow related to DefaultProxyContainerFinder and should I specify “ecf.jms.activemq.tcp.client” in the config file?
I first have to ask whether you are using the latest from HEAD or ECF 3.2. The reason I have to ask is that *since* 3.2 release the policy for handling automatic creation of containers in DefaultProxyContainerFinder has changed as per this bug: 

https://bugs.eclipse.org/bugs/show_bug.cgi?id=303979
Now (in HEAD/3.3/Helios), the default is to automatically create a container instance when discovered.
*BUT*...even if you are using HEAD....I think that there could be a bug in the ActiveMQ/JMS provider. Specifically, when containers are discovered, and the DefaultProxyContainerFinder is consulted to find a suitable container, part of what the DefaultProxyContainerFinder does (using the auto create option that was changed in 303979) is that it gets the meta-data from the discovery and looks for a *compatible* container type. By 'looks', I mean that it calls this method on container instantiatiors:
org.eclipse.ecf.core.provider.IRemoteServiceContainerInstantiator.getImportedConfigs(ContainerTypeDescription, String[])
And if the provider on the consumer responds with a non-null String array, then it has a match.
But it is quite possible that the JMS/ActiveMQ provider implementation of the getImportedConfigs method does not match with the load balancing container types (as these are 'special'), returns null, and the client container is never created/found by the DefaultProxyContainerFinder on the consumer. Does this fit with what you are seeing? BTW...what discovery mechanism are you using? And do you have evidence that it is working properly on your net?
I will be happy to look into this later today/Monday and fix it if necessary, but would it be possible to get your declarative services versions of the load balancing example so that I can test? Would you possibly be willing to contribute these projects as ECF examples? 


How is "cardinality" handled in the case of load-balanced service?
The cardinality is handled as with a 'normal' declarative service...i.e. the cardinality will determine whether the component is activated if the given reference is present (1...n) or not (0...n). 



I would appreciate any help or directions how to proceed.
I think the bug with the JMS/ActiveMQ provider I described above *may* be the issue. If you can help me diagnose and test we should get it going in short order. 


More important, is my use-case achievable with Equinox and ECF?


Yes, it is.

Scott

Back to the top