Jobs and Scheduling #
This document briefly describes how Flink schedules jobs and how it represents and tracks job status on the JobManager.
Execution resources in Flink are defined through Task Slots. Each TaskManager will have one or more task slots, each of which can run one pipeline of parallel tasks. A pipeline consists of multiple successive tasks, such as the n-th parallel instance of a MapFunction together with the n-th parallel instance of a ReduceFunction. Note that Flink often executes successive tasks concurrently: For Streaming programs, that happens in any case, but also for batch programs, it happens frequently.
The figure below illustrates that. Consider a program with a data source, a MapFunction, and a ReduceFunction. The source and MapFunction are executed with a parallelism of 4, while the ReduceFunction is executed with a parallelism of 3. A pipeline consists of the sequence Source - Map - Reduce. On a cluster with 2 TaskManagers with 3 slots each, the program will be executed as described below.
JobManager Data Structures #
During job execution, the JobManager keeps track of distributed tasks, decides when to schedule the next task (or set of tasks), and reacts to finished tasks or execution failures.
The JobManager receives the JobGraph , which is a representation of the data flow consisting of operators ( JobVertex ) and intermediate results ( IntermediateDataSet ). Each operator has properties, like the parallelism and the code that it executes. In addition, the JobGraph has a set of attached libraries, that are necessary to execute the code of the operators.
The JobManager transforms the JobGraph into an ExecutionGraph . The ExecutionGraph is a parallel version of the JobGraph: For each JobVertex, it contains an ExecutionVertex per parallel subtask. An operator with a parallelism of 100 will have one JobVertex and 100 ExecutionVertices. The ExecutionVertex tracks the state of execution of a particular subtask. All ExecutionVertices from one JobVertex are held in an ExecutionJobVertex , which tracks the status of the operator as a whole. Besides the vertices, the ExecutionGraph also contains the IntermediateResult and the IntermediateResultPartition . The former tracks the state of the IntermediateDataSet, the latter the state of each of its partitions.
Each ExecutionGraph has a job status associated with it. This job status indicates the current state of the job execution.
A Flink job is first in the created state, then switches to running and upon completion of all work it switches to finished. In case of failures, a job switches first to failing where it cancels all running tasks. If all job vertices have reached a final state and the job is not restartable, then the job transitions to failed. If the job can be restarted, then it will enter the restarting state. Once the job has been completely restarted, it will reach the created state.
In case that the user cancels the job, it will go into the cancelling state. This also entails the cancellation of all currently running tasks. Once all running tasks have reached a final state, the job transitions to the state cancelled.
Unlike the states finished, canceled and failed which denote a globally terminal state and, thus, trigger the clean up of the job, the suspended state is only locally terminal. Locally terminal means that the execution of the job has been terminated on the respective JobManager but another JobManager of the Flink cluster can retrieve the job from the persistent HA store and restart it. Consequently, a job which reaches the suspended state won’t be completely cleaned up.
During the execution of the ExecutionGraph, each parallel task goes through multiple stages, from created to finished or failed. The diagram below illustrates the states and possible transitions between them. A task may be executed multiple times (for example in the course of failure recovery). For that reason, the execution of an ExecutionVertex is tracked in an Execution . Each ExecutionVertex has a current Execution, and prior Executions.