]
Bela Ban resolved JGRP-1917.
----------------------------
Resolution: Done
Added {{remove_old_coords_on_view_change}} and {{remove_all_files_on_view_change}} to
{{FILE_PING}}. See the manual for details.
FILE_PING: options to remove zombies
------------------------------------
Key: JGRP-1917
URL:
https://issues.jboss.org/browse/JGRP-1917
Project: JGroups
Issue Type: Enhancement
Reporter: Bela Ban
Assignee: Bela Ban
Fix For: 3.6.3
In {{FILE_PING}} and subclasses ({{S3_PING}}, {{GOOGLE_PING}}), coordinators write the
files (e.g. {{A.list}} for coord {{A}}).
There's a shutdown hook that removes {{A.list}} when {{A}} crashes.
However, when a coordinator is killed by kill -9, the file {{A.list}} won't get
removed.
The problem with this is that new members will read {{A.list}} and get delayed trying to
ask {{A}} to join the cluster although {{A}}'s not alive anymore ({{B}} is and created
{{B.list}}).
Possible solution: implement a mechanism similar to JGRP-1915 where a coordinator removes
*all files* on a view change with leaving members, and then writes its file again.