[FLINK-10429] Redesign Flink Scheduling, introducing dedicated Scheduler component

• Type: New Feature
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Affects Version/s: 1.7.0
• Component/s: Runtime / Coordination
• Assignee: Gary Yao
• Reporter: Stefan Richter
• Created: 26/Sep/18 10:01
• Updated: 24/May/19 16:14

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

Issue Links

• blocks: FLINK-10240 Pluggable scheduling strategy for batch jobs
• blocks: FLINK-11298 Scheduling job in the unit of concurrent groups
• relates to: FLINK-12369 Implement a region failover strategy based on new FailoverStrategy interfaces
• relates to: FLINK-12002 Adaptive Parallelism of Job Vertex Execution
• supercedes: FLINK-10713 RestartIndividualStrategy does not restore state
• supercedes: FLINK-2250 Backtracking of intermediate results

Sub-Tasks

• Extract scheduling-related code from Executions: Open
• Introduce bulk/group-aware scheduling: Open
• Stepwise creation of the ExecutionGraph sub-structures: Open
• Add CLI command for rescaling: Open
• Enable MiniCluster tests based on schedulerNG in Flink cron build: In Progress
• Support configurable failover strategy for scheduler NG: Open
• Unify SchedulerOperations#allocateSlotsAndDeploy implementation for all scheduling strategies: Open
• Support global failure handling for DefaultScheduler (SchedulerNG): In Progress
• All task state changes should be notified to SchedulingStrategy (SchedulerNG): Open
• All partition consumable events should be notified to SchedulingStrategy (SchedulerNG): Open
• Make LazyFromSourcesSchedulingStrategy do lazy scheduling based on partition state only: Open
• Change DefaultSchedulingResultPartition to return correct partition state: Open
• Prevent vertex from being affected by outdated deployment (SchedulerNG): In Progress
• Enable ClassLoaderITCase and EventTimeWindowCheckpointingITCase to pass with scheduler NG: Open
• Enable KeyedStateCheckpointingITCase to pass with scheduler NG: Open
• Enable ZooKeeperHighAvailabilityITCase to pass with scheduler NG: Open
• Enable RegionFailoverITCase to pass with scheduler NG: Open
• Avoid to trigger failover on a non-effective task failure notification: Open
• Restore task state in new DefaultScheduler: In Progress
• RestartPipelinedRegionStrategy leverage tracked partition availability for better failover experience in DefaultScheduler: Open
• Enable BatchFineGrainedRecoveryITCase to pass with scheduler NG: Open
• Refactor SchedulingTopology to extend base topology: In Progress
• Refactor FailoverTopology to extend base topology: In Progress
• Keep only one execution topology in scheduler: In Progress
• Support building pipelined regions from base topology: In Progress
• Add a metric to show failover count regarding fine grained recovery: Open