Thanks for the report! The issue here is that a remote filter for a continuous query is loaded using peer class loading, and other classes that this remote filter depends on can be lazily loaded during its work. Loading every dependency class involves going to the node where the originating class was loading from, and asking that node to send missing classes over the network. The issues begin when this node is not in the cluster anymore, and the continuous query wasn't undeployed yet. A server sends a request for a class to a node that is not available, but wasn't kicked out of the topology yet, since a failure detection timeout hasn't elapsed yet. It leads to a NoClasDefFound exception that you observe in the logs.
The biggest issue here is that this exception triggers a failure handler that makes the whole node go down. I would expect that only one request would fail, but not the whole node. As a temporary solution you can stop relying on peer class loading for continuous queries and provide the code of remote filters to the classpath of server nodes. This way no lazy class loading will be performed over the network since they will all be available locally. Denis пт, 13 мар. 2020 г. в 20:39, VeenaMithare <v.mith...@cmcmarkets.com>: > Raised this jira : > https://issues.apache.org/jira/browse/IGNITE-12784 > > Observed in 2.7.6. Unable to easily test in 2.8.0 because of other issues. > One of them being - > > http://apache-ignite-users.70518.x6.nabble.com/2-8-0-JDBC-Thin-Client-Unable-to-load-the-tables-via-DBeaver-td31681.html > Please note this happens > > > > -- > Sent from: http://apache-ignite-developers.2346864.n4.nabble.com/ >
