-
Bug
-
Resolution: Done
-
Major
-
2.6
-
None
When flush coordinator A runs a flush for a view that excludes himself a flush coordinator transfer occurs. Member B that is first next neighbor of A according to view has to complete the flush by sending STOP_FLUSH to all.
Simplification of FLUSH JGRP-598 removed complex stop flush phase. In new design as soon as member receives STOP_FLUSH it unblocks application by sending UNBLOCK. Now, during coordinator transfer described above we first have to send a view up to application and then invoke STOP_FLUSH. If we do it other way around STOP_FLUSH will loop back and cause delivery of UNBLOCK to application prior to new view - thus causing reordering of block/unblock/view events in applications.