Github user tillrohrmann commented on the issue:

    https://github.com/apache/flink/pull/5427
  
    @EronWright, you're right that on initial submission we don't call 
`restoreLatestCheckpointedState` in the old code. With Flip-6 this will be the 
case. See #5444.
    
    The underlying assumption to make this work, though, is that a user won't 
submit a new job with the a job id to a cluster with a cluster id for which 
ZooKeeper already contains persisted checkpoints from a previous run. So either 
the cluster id or the job id must be different. 
    
    I think so far, when using the Flink client this should be the case. 
However, when generating the `JobGraph` yourself and keeping it around to 
submit it to a standalone cluster, then this assumption will break because both 
the `JobID` and the cluster id will be the same.


---

Reply via email to