Till Rohrmann created FLINK-5064:
------------------------------------
Summary: Checkpoint messages are not scoped to the leader session
ID
Key: FLINK-5064
URL: https://issues.apache.org/jira/browse/FLINK-5064
Project: Flink
Issue Type: Bug
Components: State Backends, Checkpointing
Affects Versions: 1.1.3, 1.2.0
Reporter: Till Rohrmann
Fix For: 1.2.0
The checkpoint messages ({{AbstractCheckpointMessage}}) don't implement the
{{RequiresLeaderSessionID}} interface. Thus, they are not scoped to the
leadership of a {{JobManager}} and can interfere with a new leader session.
The downside of scoping the checkpoint messages to the leader id is that
messages might get filtered out leading to resource leaks because the contained
state handle is never discarded. However, in case of a JM failure one might end
up in the same situation if there were some checkpoint messages in flight.
In order to mitigate the problem one could change the behaviour such that the
{{CheckpointResponder}} awaits a response back and in case of a negative
response or an outstanding response (timeout) it discards the state handle.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
