Problem The link and mention analysis tasks are stored with the identifier of the member that queued them, and each member only executes its own tasks. When a member never comes back with the same identifier (scale down of the cluster, identifier changed), its tasks which were not executed yet stay in the database forever: the mentioned users are never notified and the links of the document are not updated. Proposal Once a member has been absent for observation.remote.reclaimDelay, the cluster singleton node takes over its tasks queued before it left (in all the wikis) and executes them. Taking over is done in the database, so a member which comes back does not execute them a second time (a mention task is not idempotent). Remote observation disabled: no change. |
