Details
-
Bug
-
Status: Closed
-
Major
-
Resolution: Fixed
-
1.1.3, 1.2.0
-
None
Description
When deleting completed checkpoints from the ZooKeeperCompletedCheckpointStore, one first tries to delete the meta state handle from ZooKeeper and then deletes the actual checkpoint in a callback from the delete operation. This callback is executed by the ZooKeeper client's main thread which is problematic, because it blocks the ZooKeeper client. If a delete operation takes longer than it takes to complete a checkpoint, then it might even happen that delete operations of outdated checkpoints are piling up because they are effectively executed sequentially.
I propose to execute the delete operations by a dedicated Executor so that we keep the client's main thread free to do ZooKeeper related work.
