Details
-
Enhancement
-
Resolution: Done
-
Minor
-
4.1.6
-
None
Description
When the cluster is unstable but not because of the network (e.g. excessive load and long GC pauses), a suspected node can keep receiving view updates long after it was excluded from the view.
The first view triggers a warning about not being a member, but because later delta-views cannot be reconstructed, their warnings look much worse, even though they're about the same thing.
11:31:05,281 WARN [org.jgroups.protocols.pbcast.GMS] (jgroups-79,edg-perf03-47882) edg-perf03-47882: not member of view [edg-perf01-21541|6]; discarding it 11:31:16,267 WARN [org.jgroups.protocols.pbcast.GMS] (jgroups-80,edg-perf03-47882) edg-perf03-47882: failed to create view from delta-view; dropping view: java.lang.IllegalStateException: the view-id of the delta view ([edg-perf01-21541|6]) doesn't match the current view-id ([edg-perf01-21541|5]); discarding delta view [edg-perf01-21541|7], ref-view=[edg-perf01-21541|6], left=[edg-perf06-47720]
Maybe GMS could keep track of when it receives a view in which it is not a member and stop processing delta views from that node?
