Uploaded image for project: 'Flink'
  1. Flink
  2. FLINK-10429

Redesign Flink Scheduling, introducing dedicated Scheduler component

    XMLWordPrintableJSON

    Details

    • Type: New Feature
    • Status: Open
    • Priority: Major
    • Resolution: Unresolved
    • Affects Version/s: 1.7.0
    • Fix Version/s: None
    • Component/s: Runtime / Coordination
    • Labels:
      None

      Description

      This epic tracks the redesign of scheduling in Flink. Scheduling is currently a concern that is scattered across different components, mainly the ExecutionGraph/Execution and the SlotPool. Scheduling also happens only on the granularity of individual tasks, which make holistic scheduling strategies hard to implement. In this epic we aim to introduce a dedicated Scheduler component that can support use-case like auto-scaling, local-recovery, and resource optimized batch.

      The design for this feature is developed here: https://docs.google.com/document/d/1q7NOqt05HIN-PlKEEPB36JiuU1Iu9fnxxVGJzylhsxU/edit?usp=sharing

        Attachments

          Issue Links

          blocks

          New Feature - A new feature of the product, which has yet to be developed. FLINK-10240 Pluggable scheduling strategy for batch jobs

          • Major - Major loss of function.
          • Open

          Improvement - An improvement or enhancement to an existing feature or task. FLINK-11298 Scheduling job in the unit of concurrent groups

          • Major - Major loss of function.
          • Open
          relates to

          Sub-task - The sub-task of the issue FLINK-12369 Implement a region failover strategy based on new FailoverStrategy interfaces

          • Major - Major loss of function.
          • Closed

          Improvement - An improvement or enhancement to an existing feature or task. FLINK-12002 Adaptive Parallelism of Job Vertex Execution

          • Major - Major loss of function.
          • Open
          supercedes

          Bug - A problem which impairs or prevents the functions of the product. FLINK-10713 RestartIndividualStrategy does not restore state

          • Critical - Crashes, loss of data, severe memory leak.
          • Open

          New Feature - A new feature of the product, which has yet to be developed. FLINK-2250 Backtracking of intermediate results

          • Major - Major loss of function.
          • Closed
          1.
          		 Make access to ExecutionGraph single threaded from JobMaster main thread Sub-task Closed Stefan Richter
          2.
          		 Extract scheduling-related code from Executions Sub-task Open Stefan Richter  
          3.
          		 Extract scheduling-related code from SlotPool Sub-task Closed Stefan Richter
          4.
          		 Introduce bulk/group-aware scheduling Sub-task Open Stefan Richter  
          5.
          		 Stepwise creation of the ExecutionGraph sub-structures Sub-task Open Stefan Richter  
          6.
          		 Introduce SchedulingStrategy interface Sub-task Closed shuai.xu
          7.
          		 Implement Eager Scheduling Strategy Sub-task Closed shuai.xu
          8.
          		 Implement Lazy Scheduling Strategy Sub-task Closed BoWang
          9.
          		 Remove getExecutionGraph() from JobMaster Sub-task Closed Gary Yao
          10.
          		 Introduce Scheduler interface and adapt ExecutionGraph to it Sub-task Closed Gary Yao
          11.
          		 Temporarily disable CLI command for rescaling Sub-task Closed Gary Yao
          12.
          		 Add CLI command for rescaling Sub-task Open Unassigned  
          13.
          		 Implement ExecutionSlotAllocator Sub-task Closed shuai.xu
          14.
          		 Implement ExecutionFailureHandler Sub-task Closed Zhu Zhu
          15.
          		 Implement ExecutionGraph to SchedulingTopology Adapter Sub-task Closed BoWang
          16.
          		 Add SchedulerNG Stub Implementation Sub-task Closed Gary Yao
          17.
          		 Implement DefaultScheduler stub Sub-task Closed Gary Yao
          18.
          		 Add getVertex/ResultPartitionOrThrow(ExecutionVertexID/IntermediateResultPartitionID) to SchedulingTopology Sub-task Closed BoWang
          19.
          		 Implement FixedDelayRestartBackoffTimeStrategy Sub-task Closed BoWang
          20.
          		 Implement FailureRateRestartBackoffTimeStrategy Sub-task Closed BoWang
          21.
          		 Implement RestartBackoffTimeStrategyFactoryLoader Sub-task Closed Zhu Zhu
          22.
          		 Implement ExecutionGraph to InputsLocationsRetriever Adapter Sub-task Closed shuai.xu
          23.
          		 Facilitate enabling new Scheduler in MiniCluster Tests Sub-task Closed Zhu Zhu
          24.
          		 Enable NG scheduler testing in per-commit tests Sub-task Closed Zhu Zhu
          25.
          		 Rewrite DefaultExecutionSlotAllocator to use SlotProviderStrategy Sub-task Closed Gary Yao
          26.
          		 Remove the description of restart strategy customization Sub-task Closed Zhu Zhu
          27.
          		 Support configurable failover strategy for scheduler NG Sub-task Closed Zhu Zhu  
          28.
          		 Unify SchedulerOperations#allocateSlotsAndDeploy implementation for all scheduling strategies Sub-task Open Unassigned  
          29.
          		 Support global failure handling for DefaultScheduler (SchedulerNG) Sub-task Closed Zhu Zhu
          30.
          		 All task state changes should be notified to SchedulingStrategy (SchedulerNG) Sub-task Open Unassigned  
          31.
          		 All partition consumable events should be notified to SchedulingStrategy (SchedulerNG) Sub-task Open Unassigned  
          32.
          		 Make LazyFromSourcesSchedulingStrategy do lazy scheduling based on partition state only Sub-task Open Unassigned  
          33.
          		 Use correct RestartBackoffTimeStrategy in new DefaultScheduler Sub-task Closed Gary Yao
          34.
          		 Annotate all MiniCluster tests in flink-runtime with AlsoRunWithSchedulerNG Sub-task Closed Zhu Zhu
          35.
          		 Use NoResourceAvailableException to wrap TimeoutException on slot allocation (Scheduler NG) Sub-task Closed Zhu Zhu
          36.
          		 LazyFromSourcesSchedulingStrategy should be able to restart terminated tasks Sub-task Closed Zhu Zhu
          37.
          		 Add unit tests for DefaultScheduler to test concurrent failover behavior Sub-task Closed Gary Yao
          38.
          		 Remove legacy RestartPipelinedRegionStrategy Sub-task Closed Zhu Zhu
          39.
          		 Get ExecutionVertexID from ExecutionVertex rather than creating new instances Sub-task Closed Zhu Zhu
          40.
          		 Introduce a unified topology interface Sub-task Resolved Zhu Zhu
          41.
          		 Change DefaultSchedulingResultPartition to return correct partition state Sub-task Closed Zhu Zhu
          42.
          		 Reset vertices right after they transition to terminated states Sub-task Closed Zhu Zhu
          43.
          		 Prevent vertex from being affected by outdated deployment (SchedulerNG) Sub-task Closed Zhu Zhu
          44.
          		 Annotate MiniCluster tests in core modules with AlsoRunWithSchedulerNG Sub-task Resolved Zhu Zhu
          45.
          		 Annotate MiniCluster tests in flink-tests with AlsoRunWithSchedulerNG Sub-task Resolved Zhu Zhu
          46.
          		 Enable ClassLoaderITCase to pass with scheduler NG Sub-task Closed Zhu Zhu
          47.
          		 Enable KeyedStateCheckpointingITCase to pass with scheduler NG Sub-task Closed Gary Yao
          48.
          		 Enable ZooKeeperHighAvailabilityITCase to pass with scheduler NG Sub-task Closed Zhu Zhu
          49.
          		 Enable RegionFailoverITCase to pass with scheduler NG Sub-task Closed Gary Yao
          50.
          		 Avoid to notify ineffective state updates to scheduler Sub-task Closed Zhu Zhu
          51.
          		 Restore task state in new DefaultScheduler Sub-task Closed Zhu Zhu
          52.
          		 RestartPipelinedRegionStrategy leverage tracked partition availability for better failover experience in DefaultScheduler Sub-task Closed Zhu Zhu
          53.
          		 Enable BatchFineGrainedRecoveryITCase to pass with scheduler NG Sub-task Closed Zhu Zhu
          54.
          		 Refactor SchedulingTopology to extend base topology Sub-task Resolved Zhu Zhu
          55.
          		 Refactor FailoverTopology to extend base topology Sub-task Resolved Zhu Zhu
          56.
          		 Keep only one execution topology in scheduler Sub-task Resolved Zhu Zhu
          57.
          		 Support building pipelined regions from base topology Sub-task Resolved Zhu Zhu
          58.
          		 Add a metric to show failover count regarding fine grained recovery Sub-task Closed Zhu Zhu
          59.
          		 Simplify params of Execution#processFail Sub-task Open Unassigned  
          60.
          		 Handle schedule mode LAZY_FROM_SOURCES_WITH_BATCH_SLOT_REQUEST correctly in DefaultScheduler Sub-task Closed Gary Yao
          61.
          		 Enable KafkaConsumerTestBase#runFailOnNoBrokerTest to pass with new DefaultScheduler Sub-task Closed Gary Yao
          62.
          		 Enable YARNHighAvailabilityITCase to pass with new DefaultScheduler Sub-task Closed Gary Yao  
          63.
          		 Enable AbstractTaskManagerProcessFailureRecoveryTest to pass with new DefaultScheduler Sub-task Closed Zhu Zhu  
          64.
          		 Introduce full restarts failover strategy for NG scheduler Sub-task Closed Zhu Zhu
          65.
          		 Avoid leaking instance of DefaultScheduler before object is constructed Sub-task Closed Gary Yao
          66.
          		 Enable 'Queryable state (rocksdb) with TM restart' E2E test to pass with new DefaultScheduler Sub-task Closed Gary Yao
          67.
          		 Enable 'Streaming File Sink end-to-end test' to pass with new DefaultScheduler Sub-task Closed Gary Yao
          68.
          		 Enable 'Streaming bucketing end-to-end test' to pass with new DefaultScheduler Sub-task Closed Gary Yao
          69.
          		 Avoid leaking unassigned Slot in DefaultScheduler when Deployment is outdated Sub-task Closed Gary Yao
          70.
          		 Let tasks in a batch get scheduled in topological order and subtaskIndex ascending pattern Sub-task Closed Zhu Zhu
          71.
          		 Set default value of config option jobmanager.scheduler to "ng" Sub-task Closed Gary Yao
          72.
          		 SchedulerBase should only log the RestartStrategy in legacy scheduling mode Sub-task Closed Gary Yao

            Activity

              People

              • Assignee:
                gjy Gary Yao
                Reporter:
                srichter Stefan Richter
              • Votes:
                3 Vote for this issue
                Watchers:
                45 Start watching this issue

                Dates

                • Created:
                  Updated:

                  Time Tracking

                  Estimated:
                  Original Estimate - Not Specified
                  Not Specified
                  Remaining:
                  Remaining Estimate - 0h
                  0h
                  Logged:
                  Time Spent - 22.5h
                  22.5h