Package org.apache.ignite.compute

Interface ComputeJob

  • All Superinterfaces:
    Serializable
    All Known Implementing Classes:
    ComputeJobAdapter, ComputeJobContinuationAdapter
    public interface ComputeJob
extends Serializable
    Defines executable unit for ComputeTask.

    Description

    Grid job is an executable unit of ComputeTask. Grid task gets split into jobs when ComputeTask.map(List, Object) method is called. This method returns all jobs for the task mapped to their corresponding grid nodes for execution. Grid will then serialize this jobs and send them to requested nodes for execution. When a node receives a request to execute a job, the following sequence of events takes place:
    1. If collision SPI is defined, then job gets put on waiting list which is passed to underlying CollisionSpi SPI. Otherwise job will be submitted to the executor service responsible for job execution immediately upon arrival.
    2. If collision SPI is configured, then it will decide one of the following scheduling policies:
      • Job will be kept on waiting list. In this case, job will not get a chance to execute until next time the Collision SPI is called.
      • Job will be moved to active list. In this case system will proceed with job execution.
      • Job will be rejected. In this case the ComputeJobResult passed into ComputeTask.result(ComputeJobResult, List) method will contain ComputeExecutionRejectedException exception. If you are using any of the task adapters shipped with Ignite, then job will be failed over automatically for execution on another node.
    3. For activated jobs, an instance of distributed task session (see ComputeTaskSession) will be injected.
    4. System will execute the job by calling execute() method.
    5. If job gets cancelled while executing then cancel() method will be called. Note that just like with Thread.interrupt() method, grid job cancellation serves as a hint that a job should stop executing or exhibit some other user defined behavior. Generally it is up to a job to decide whether it wants to react to cancellation or ignore it. Job cancellation can happen for several reasons:
      • Collision SPI cancelled an active job.
      • Parent task has completed without waiting for this job's result.
      • User cancelled task by calling IgniteFuture.cancel() method.
    6. Once job execution is complete, the return value will be sent back to parent task and will be passed into ComputeTask.result(ComputeJobResult, List) method via ComputeJobResult instance. If job execution resulted in a checked exception, then ComputeJobResult.getException() method will contain that exception. If job execution threw a runtime exception or error, then it will be wrapped into ComputeUserUndeclaredException exception.

    Resource Injection

    Grid job implementation can be injected using IoC (dependency injection) with ignite resources. Both, field and method based injection are supported. The following ignite resources can be injected: Refer to corresponding resource documentation for more information.

    ComputeJobAdapter

    Ignite comes with convenience ComputeJobAdapter adapter that provides default empty implementation for cancel() method and also allows user to set and get job argument, if there is one.

    Distributed Session Attributes

    Jobs can communicate with parent task and with other job siblings from the same task by setting session attributes (see ComputeTaskSession). Other jobs can wait for an attribute to be set either synchronously or asynchronously. Such functionality allows jobs to synchronize their execution with other jobs at any point and can be useful when other jobs within task need to be made aware of certain event or state change that occurred during job execution.

    Distributed task session can be injected into ComputeJob implementation using @TaskSessionResource annotation. Both, field and method based injections are supported. Refer to ComputeTaskSession documentation for more information on session functionality.

    Saving Checkpoints

    Long running jobs may wish to save intermediate checkpoints to protect themselves from failures. There are three checkpoint management methods: Jobs that utilize checkpoint functionality should attempt to load a check point at the beginning of execution. If a non-null value is returned, then job can continue from where it failed last time, otherwise it would start from scratch. Throughout it's execution job should periodically save its intermediate state to avoid starting from scratch in case of a failure.

    • Method Summary

      All Methods Instance Methods Abstract Methods 
      Modifier and Type Method Description
      void cancel()
      This method is called when system detects that completion of this job can no longer alter the overall outcome (for example, when parent task has already reduced the results).
      Object execute()
      Executes this job.