Regression in messaging cluster tests with network failures ----------------------------------------------------------- Key: WFWIP-13 URL: https://issues.jboss.org/browse/WFWIP-13 Project: WildFly WIP Issue Type: Bug Components: Artemis Reporter: Erich Duda Assignee: Yong Hao Gao Priority: Blocker Labels: feature-branch-blocker *Scenario* * There are two Artemis brokers configured to form cluster * There is a producer sending messages to broker 1 and receiver receiving messages from broker 2 * Between the brokers there is a proxy which simulates network failure * The proxy is several times stopped and restarted to simulate the network failure * The test expects that all messages sent to broker 1 will be received by receiver from broker 2 (despite the network failures) *Reality:* After the proxy is stopped and restarted, the cluster is not able to form again. Both brokers try to reconnect to their opposites but with no luck. *Customer scenario:* Messaging cluster is not able to recover after network failures. *Investigation of issue* I investigated why brokers are not able to reconnect and I found out that always when they try to reconnect, they give it up because there is no topology record for {{nodeId}} where they try to connect. So the re-connection attempt ends here \[1\]. I compared the behavior with Artemis 1.x and I found out that Artemis 2.x removes the topology member when connection failure is detected, but Artemis 1.x doesn't. When I commented the line \[2\] it fixed the issue. This line is not present in 1.x. \[1\] https://github.com/apache/activemq-artemis/blob/b66d0f7ac40001cce14ca7146... \[2\] https://github.com/apache/activemq-artemis/blob/b66d0f7ac40001cce14ca7146...