Pierre Villard created NIFI-16563:
-------------------------------------
Summary: Cluster merge of Controller Service state ignores nodes
that have not started transitioning
Key: NIFI-16563
URL: https://issues.apache.org/jira/browse/NIFI-16563
Project: Apache NiFi
Issue Type: Bug
Reporter: Pierre Villard
Assignee: Pierre Villard
When a Controller Service is enabled or disabled in a cluster, each node
schedules the state change asynchronously and responds before it starts.
ControllerServiceEntityMerger only reports a transitional state when a node is
ENABLING or DISABLING. If one node has already reached DISABLED while another
still reports ENABLED, the merged response reflects a single node.
ClusterReplicationComponentLifecycle can then consider the transition complete
too early. For example, a Parameter Context update can be applied while a
service is still enabled on one node. That node rejects the request, the
coordinator asks it to reconnect, and the update fails. Seen intermittently in
ClusteredParameterContextIT.testProcessorRestartedAfterLongDependentServiceValidationPeriod.
The merger now reports DISABLING when nodes disagree between ENABLED and
DISABLED.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
