Till Rohrmann created FLINK-5065:
------------------------------------
Summary: Resource leaks in case of lost checkpoint messages
Key: FLINK-5065
URL: https://issues.apache.org/jira/browse/FLINK-5065
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
In case of lost checkpoint messages ({{AbstractCheckpointMessage}}) there won't
be a clean up of the contained state handle done. This can happen in case of a
JM crash if the TM had just sent a checkpoint message to the JM. Upon recovery,
the JM won't notice that there was a checkpoint message which it had to handle.
One way to mitigate the problem is to let the TM expect a response from the JM
and if the JM failed to respond in time, to discard the state handle of the
checkpoint message. Moreover, the JM could register in a fault tolerant store
that it started a pending checkpoint. The entry should contain enough
information (e.g. path for {{FileSystemStateBackend}}) to clean up a pending
checkpoint with potentially lost checkpoint messages upon recovery of the JM.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
