shashank created CAMEL-27337:
--------------------------------

             Summary: camel-jgroups-raft - the cluster views of the namespaces 
of a node do not share the RaftHandle correctly: a late namespace never leads, 
stopping one disconnects the others, and without a configured handle a second 
namespace cannot start
                 Key: CAMEL-27337
                 URL: https://issues.apache.org/jira/browse/CAMEL-27337
             Project: Camel
          Issue Type: Bug
          Components: camel-jgroups
            Reporter: shashank
            Assignee: shashank


h3. What happens

{{JGroupsRaftClusterService}} gives one view per namespace (master and 
clustered routes). A node is one Raft member (one raft id), so the views have 
to share one {{RaftHandle}}, but each view manages the handle as if it were its 
own:

# A view learns its role only from role change callbacks. Its {{doStart}} 
registers the listener and calls {{channel().connect}}, which does nothing once 
the shared channel is connected, and never reads the current role. A namespace 
whose view starts after the node was elected (a master route added or restarted 
at runtime, or started lazily) keeps {{isMaster=false}}, and as the other nodes 
are followers no node runs it until the Raft leader changes.
# When the last user of a namespace releases its view, {{doStop}} disconnects 
the shared channel and closes its log. jgroups-raft steps the node down, so 
every other namespace loses the leadership and the node stays out of Raft; in a 
2 node cluster no namespace has a leader anywhere.
# Without a configured {{raftHandle}} (the default), each view creates its own 
{{JChannel}} and {{RaftHandle}} with the service's single raft id. {{NO_DUPES}} 
rejects the second member, and after 10 tries of 5 s the second route fails to 
start: {{SecurityException: join of A rejected as it would create a view with 
duplicate members}}.

h3. How to reproduce

One node ({{raft.xml}} of the module's tests), two master routes 
{{master:ns1:timer:...}} and {{master:ns2:timer:...}}: add the second route 
after the node became leader (it never runs), or stop the first route (the 
second one stops running), or use no {{raftHandle}} on the service (the 
CamelContext does not start).

Found with a TLA+ model of the cluster views.

h3. Fix

The service owns the node's handle: the configured one or one it creates, 
connected when the first view starts and disconnected when the last view stops 
(reference counted, as {{JGroupsRaftEndpoint}} and the component do for their 
handles, CAMEL-25883). One role listener on the handle forwards the role 
changes to the views; a view reads the current role after it connected, and 
ignores the role changes still notified to it once it is stopped.

_Filed with Claude Code on behalf of allthingssecurity._



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to