XMLWordPrintableJSON

    Details

    • Type: Sub-task
    • Status: Resolved
    • Priority: Major
    • Resolution: Fixed
    • Affects Version/s: 0.10.0
    • Fix Version/s: 0.10.0
    • Component/s: Runtime / Coordination
    • Labels:
      None

      Description

      tl;dr Persist JobGraphs in state backend and coordinate reference to state handle via ZooKeeper.

      Problem: When running multiple JobManagers in high availability mode, the leading job manager looses all running jobs when it fails. After a new leading job manager is elected, it is not possible to recover any previously running jobs.

      Solution: The leading job manager, which receives the job graph writes 1) the job graph to a state backend, and 2) a reference to the respective state handle to ZooKeeper. In general, job graphs can become large (multiple MBs, because they include closures etc.). ZooKeeper is not designed for data of this size. The level of indirection via the reference to the state backend keeps the data in ZooKeeper small.

      Proposed ZooKeeper layout:

      /flink (default)
      +- currentJobs
      +- job id i
      +- state handle reference of job graph i

      The 'currentJobs' node needs to be persistent to allow recovery of jobs between job managers. The currentJobs node needs to satisfy the following invariant: There is a reference to a job graph with id i IFF the respective job graph needs to be recovered by a newly elected job manager leader.

      With this in place, jobs will be recovered from their initial state (as if resubmitted). The next step is to backup the runtime state handles of checkpoints in a similar manner.

      This work will be based on Till Rohrmann's implementation of FLINK-2291. The leader election service notifies the job manager about granted/revoked leadership. This notification happens via Akka and thus serially per job manager, but results in eventually consistent state between job managers. For some snapshots of time it is possible to have a new leader granted leadership, before the old one has been revoked its leadership.

      Till Rohrmann, can you confirm that leadership does not guarantee mutually exclusive access to the shared 'currentJobs' state?

      For example, the following can happen:

      • JM 1 is leader, JM 2 is standby
      • JOB i is running (and hence /flink/currentJobs/i exists)
      • ZK notifies leader election service (LES) of JM 1 and JM 2
      • LES 2 immediately notifies JM 2 about granted leadership, but LES 1 notification revoking leadership takes longer
      • JOB i finishes (TMs don't notice leadership change yet) and JM 1 receives final JobStatusChange
      • JM 2 resubmits the job /flink/currentJobs/i
      • JM 1 removes /flink/currentJobs/i, because it is now finished
        => inconsistent state (wrt the specified invariant above)

      If it is indeed a problem, we can circumvent this with a Curator recipe for shared locks to coordinate the access to currentJobs. The lock needs to be acquired on leadership.

      Minimum required tests:

      • Unit tests for job graph serialization and writing to state backend and ZooKeeper with expected nodes
      • Unit tests for job submission to job manager in leader/non-leader state
      • Unit tests for leadership granting/revoking and job submission/restarting interleavings
      • Process failure integration tests with single and multiple running jobs

        Attachments

          Issue Links

            Activity

              People

              • Assignee:
                uce Ufuk Celebi
                Reporter:
                uce Ufuk Celebi
              • Votes:
                0 Vote for this issue
                Watchers:
                5 Start watching this issue

                Dates

                • Created:
                  Updated:
                  Resolved: