Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
1.5.0
-
None
-
None
Description
RaftConsensus maintains a failed_elections_since_stable_leader_ counter used to cause elections to back off on failure. However, if a replica is partitioned for a while such that it calls some failed pre-elections, and then gets re-connected to the cluster such that the leader never changed, the counter will remain high. It only resets on an actual leader change. This means that if the leader did actually fail in the future, the failure would not be detected for a potentially long time.
Instead we should reset the counter on any successful update from the leader.