Recovery job progress from job master failures
This documentation is for an unreleased version of Apache Flink. We recommend you use the latest stable version.

Batch jobs progress recovery from job master failures #

Background #

Previously, if the JobMaster fails and is terminated, one of the following two situations will occur:

  • If high availability (HA) is disabled, the job will fail.
  • If HA is enabled, a JobMaster failover will happen and the job will be restarted. Streaming jobs can resume from the latest successful checkpoints. Batch jobs, however, do not have checkpoints and have to start over from the beginning, losing all previously made progress. This represents a significant regression for long-running batch jobs.

To address this issue, a batch job recovery mechanism is introduced to enable batch jobs to recover as much progress as possible after a JobMaster failover, avoiding the need to rerun tasks that have already been finished.

To implement this feature, a JobEventStore component is introduced to record state change events of the JobMaster (such as ExecutionGraph, OperatorCoordinator, etc.) to an external filesystem. During the crash and subsequent restart of the JobMaster, TaskManagers will retain the intermediate result data produced by the job and attempt to reconnect continuously. Once the JobMaster restarts, it will re-establish connections with TaskManagers and recover the job state based on the retained intermediate results and the events previously recorded in the JobEventStore, thereby resuming the job’s execution progress.

Usage #

This section explains how to enable recovery of batch jobs from JobMaster failures, how to tune it, and how to develop sources to work with batch jobs progress recovery.

How to enable batch jobs progress recovery from job master failures #

  • Enable cluster high availability:

    To enable the recovery of batch jobs from JobMaster failures, it is essential to first ensure that cluster high availability (HA) is enabled. Flink supports HA services backed by ZooKeeper or Kubernetes. More details of the configuration can be found in the High Availability page.

  • Configure execution.batch.job-recovery.enabled: true

Note that currently only Adaptive Batch Scheduler supports this feature. And Flink batch jobs will use this scheduler by default unless another scheduler is explicitly configured.

Optimization #

To enable batch jobs to recover as much progress as possible after a JobMaster failover, and avoid rerunning tasks that have already been finished, you can configure the following options for optimization:

  • execution.batch.job-recovery.snapshot.min-pause: This setting determines the minimum pause time allowed between snapshots for the OperatorCoordinator and ShuffleMaster. This parameter could be adjusted based on the expected I/O load of your cluster and the tolerable amount of state regression. Reduce this interval if smaller state regressions are preferred and a higher I/O load is acceptable.
  • execution.batch.job-recovery.previous-worker.recovery.timeout: This setting determines the timeout duration allowed for Shuffle workers to reconnect. During the recovery process, Flink requests the retained intermediate result data information from the Shuffle Master. If the timeout is reached, Flink will use all the acquired intermediate result data to recover the state.
  • job-event.store.write-buffer.flush-interval: This setting determines the flush interval for the JobEventStore’s write buffers.
  • job-event.store.write-buffer.size: This setting determines the write buffer size in the JobEventStore. When the buffer is full, its contents are flushed to the external filesystem.

Enable batch jobs progress recovery for sources #

Currently, only the new source (FLIP-27) supports progress recovery for batch jobs. To achieve this functionality, the SplitEnumerator of the new source (FLIP-27) must be able to take state snapshots in batch processing scenarios (where the checkpointId is set to -1) and implement the SupportsBatchSnapshot interface. This allows it to recover to the progress before the job master failure. Otherwise, to ensure data accuracy, one of the following two situations will occur after a job master failover:

  1. If not all tasks of this source are finished, we will reset and re-run all these tasks.
  2. If all tasks of this source are finished, no additional action is required, and the job can continue to run. However, if any of these tasks need to be restarted at some point in the future (for example, due to a PartitionNotFound exception), then all subtasks of this source will need to be reset and rerun.

Limitations #

  • Only working with the new source (FLIP-27): Since the legacy source has been deprecated, this feature only supports the new source.
  • Exclusive to the Adaptive Batch Scheduler: Currently, only the Adaptive Batch Scheduler supports the recovery of batch jobs after a JobMaster failover. As a result, the feature inherits all the limitations of the Adaptive Batch Scheduler.
  • Not working when using remote shuffle services.