[CONTROLLER-1927] Transaction can become stuck in COMMIT_PENDING when a node flaps leader -> follower -> leader

Relates — Mon, 27 Apr 2020 07:35:55 +0000

Normally an entry application is as follows:

leader sends an append entry off to persistence and replicates it to followers
leaders creates its ClientRequestTracker
when the entry is done with persistence and replication leader moves its commit index
part of moving the commit index is sending an ApplyState message which finalizes the entry application in the DataTree
The ApplyState determines if a ClientRequestTracker is present and adds an identifier to the ApplyState message if it is.
- This determines the way in which the finalize of the entry application happens in the DataTree.
- If it is present the entry is applied as if it originated on the leader,
- if it is not present it is applied as if the node is a follower.

The problem is when the leader flaps in a leader -> follower -> leader transition after 2. and before 4.. This would mean that the new leader no longer has the ClientRequestTracker which was created in the previous leader state, which means
that when it starts with 5. It will create the ApplyState without an identifier and the entry finishes up the application as if the node is a follower.

This means that it will be applied without finishCommit which means that the transaction will be forever stuck in COMMIT_PENDING state until the node would be restarted.

OpenDaylight JIRA

[CONTROLLER-1927] Transaction can become stuck in COMMIT_PENDING when a node flaps leader -> follower -> leader