JGRP-348 describes an edge case that causes interruption of unicast communication between two separate unmerged JGroups network islands. However, a fix for JGRP-348 introduced even a bigger edge case of the same problem described in JGRP-659.

For 2.6.3 release we will revert changes made in JGRP-348. To handle an edge case described in JGRP-348 we recommend use of FD_ALL failure detection protocol. JGRP-659 will be solved properly in 2.7 to handle both edge cases out of the box.