Discovery does not recognise new instances or removed instances

[sigrist]

Hi all,

I'm starting to use Zookeeper for service discovery with Spring Cloud. I created a simple application to demonstrate my issue. My client does not recognise when new instances are available or, even worse, when an instance is not available. The code for this example is https://github.com/sigrist/zookeeper-sandbox

I created a simple HTTP service that sends a Hello. I build a client service that uses this 'Hello Service'. If I started 2 or more services before the client, I can see that the load balancer works as expected, calling all the available instances. But if I shutdown one of the servers, the client still tries to connect to it, failing and showing the Default value for a failed request (using Netflix Hystrix).

If I started new server instances, the client does not recognise then, using only the available at boot time.

When I started new servers, I can see in the Client log that a new instance is available, but the server list is not updated:

2015-11-02 12:36:36.314  INFO 35764 --- [-ServiceCache-0] .w.DependencyStateChangeListenerRegistry : Service cache state change for '/zookeeper-sandbox' instances, current service state: CONNECTED

Also, when I shutdown one server instance, a log in the client shows a new event:

2015-11-02 12:37:29.646  INFO 35764 --- [-ServiceCache-0] .w.DependencyStateChangeListenerRegistry : Service cache state change for '/zookeeper-sandbox' instances, current service state: CONNECTED

My question is, how to update the server list in the client? Is there some configuration that I'm missing?

Thanks
Sigrist

[spencergibb]

I haven't tried your project yet, but the sample recognizes new and removed instances.

[spencergibb]

So the issue is that DependenciesBasedLoadBalancer does not extend DynamicServerListLoadBalancer, just BaseLoadBalancer which doesn't update the list of instances from Zookeeper. The workaround is to not use the dependencies for now. /cc @marcingrzejszczak

[spencergibb]

Thanks @sigrist for the report and @marcingrzejszczak for the fix!