Skip to content

DAGScheduler

Note

The introduction that follows was highly influenced by the scaladoc of org.apache.spark.scheduler.DAGScheduler. As DAGScheduler is a private class it does not appear in the official API documentation. You are strongly encouraged to read the sources and only then read this and the related pages afterwards.

Introduction

DAGScheduler is the scheduling layer of Apache Spark that implements stage-oriented scheduling using Jobs and Stages.

DAGScheduler transforms a logical execution plan (RDD lineage of dependencies built using RDD transformations) to a physical execution plan (using stages).

DAGScheduler Transforming RDD Lineage Into Stage DAG

After an action has been called on an RDD, SparkContext hands over a logical plan to DAGScheduler that it in turn translates to a set of stages that are submitted as TaskSets for execution.

Executing action leads to new ResultStage and ActiveJob in DAGScheduler

DAGScheduler works solely on the driver and is created as part of SparkContext's initialization (right after TaskScheduler and SchedulerBackend are ready).

DAGScheduler as created by SparkContext with other services

DAGScheduler does three things in Spark:

  • Computes an execution DAG (DAG of stages) for a job
  • Determines the preferred locations to run each task on
  • Handles failures due to shuffle output files being lost

DAGScheduler computes a directed acyclic graph (DAG) of stages for each job, keeps track of which RDDs and stage outputs are materialized, and finds a minimal schedule to run jobs. It then submits stages to TaskScheduler.

DAGScheduler.submitJob

In addition to coming up with the execution DAG, DAGScheduler also determines the preferred locations to run each task on, based on the current cache status, and passes the information to TaskScheduler.

DAGScheduler tracks which rdd/spark-rdd-caching.md[RDDs are cached (or persisted)] to avoid "recomputing" them, i.e. redoing the map side of a shuffle. DAGScheduler remembers what ShuffleMapStage.md[ShuffleMapStage]s have already produced output files (that are stored in BlockManagers).

DAGScheduler is only interested in cache location coordinates, i.e. host and executor id, per partition of a RDD.

Furthermore, it handles failures due to shuffle output files being lost, in which case old stages may need to be resubmitted. Failures within a stage that are not caused by shuffle file loss are handled by the TaskScheduler itself, which will retry each task a small number of times before cancelling the whole stage.

DAGScheduler uses an event queue architecture in which a thread can post DAGSchedulerEvent events, e.g. a new job or stage being submitted, that DAGScheduler reads and executes sequentially. See the section <>.

DAGScheduler runs stages in topological order.

DAGScheduler uses SparkContext, TaskScheduler, LiveListenerBus.md[], MapOutputTracker.md[MapOutputTracker] and storage:BlockManager.md[BlockManager] for its services. However, at the very minimum, DAGScheduler takes a SparkContext only (and requests SparkContext for the other services).

When DAGScheduler schedules a job as a result of rdd/index.md#actions[executing an action on a RDD] or calling SparkContext.runJob() method directly, it spawns parallel tasks to compute (partial) results per partition.

Creating Instance

DAGScheduler takes the following to be created:

DAGScheduler is created when SparkContext is created.

While being created, DAGScheduler requests the TaskScheduler to associate itself with and requests DAGScheduler Event Bus to start accepting events.

DAGSchedulerSource

DAGScheduler uses DAGSchedulerSource for performance metrics.

DAGScheduler Event Bus

DAGScheduler uses an event bus to process scheduling events on a separate thread (one by one and asynchronously).

DAGScheduler requests the event bus to start right when created and stops it when requested to stop.

DAGScheduler defines event-posting methods for posting DAGSchedulerEvent events to the event bus.

TaskScheduler

DAGScheduler is given a TaskScheduler when created.

TaskScheduler is used for the following:

Running Job

runJob[T, U](
  rdd: RDD[T],
  func: (TaskContext, Iterator[T]) => U,
  partitions: Seq[Int],
  callSite: CallSite,
  resultHandler: (Int, U) => Unit,
  properties: Properties): Unit

runJob submits a job and waits until a result is available.

runJob prints out the following INFO message to the logs when the job has finished successfully:

Job [jobId] finished: [callSite], took [time] s

runJob prints out the following INFO message to the logs when the job has failed:

Job [jobId] failed: [callSite], took [time] s

runJob is used when SparkContext is requested to run a job.

Submitting Job

submitJob[T, U](
  rdd: RDD[T],
  func: (TaskContext, Iterator[T]) => U,
  partitions: Seq[Int],
  callSite: CallSite,
  resultHandler: (Int, U) => Unit,
  properties: Properties): JobWaiter[U]

submitJob increments the nextJobId internal counter.

submitJob creates a JobWaiter for the (number of) partitions and the given resultHandler function.

submitJob requests the DAGSchedulerEventProcessLoop to post a JobSubmitted.

In the end, submitJob returns the JobWaiter.

For empty partitions (no partitions to compute), submitJob requests the LiveListenerBus to post a SparkListenerJobStart and SparkListenerJobEnd (with JobSucceeded result marker) events and returns a JobWaiter with no tasks to wait for.

submitJob throws an IllegalArgumentException when the partitions indices are not among the partitions of the given RDD:

Attempting to access a non-existent partition: [p]. Total number of partitions: [maxPartitions]

submitJob is used when:

Partition Placement Preferences

DAGScheduler keeps track of block locations per RDD and partition.

DAGScheduler uses TaskLocation that includes a host name and an executor id on that host (as ExecutorCacheTaskLocation).

The keys are RDDs (their ids) and the values are arrays indexed by partition numbers.

Each entry is a set of block locations where a RDD partition is cached, i.e. the BlockManagers of the blocks.

Initialized empty when DAGScheduler is created.

Used when DAGScheduler is requested for the locations of the cache blocks of a RDD.

ActiveJobs

DAGScheduler tracks ActiveJobs:

DAGScheduler uses ActiveJobs registry when requested to handle JobGroupCancelled or TaskCompletion events, to cleanUpAfterSchedulerStop and to abort a stage.

The number of ActiveJobs is available using job.activeJobs performance metric.

Creating ResultStage for RDD

createResultStage(
  rdd: RDD[_],
  func: (TaskContext, Iterator[_]) => _,
  partitions: Array[Int],
  jobId: Int,
  callSite: CallSite): ResultStage

createResultStage...FIXME

createResultStage is used when DAGScheduler is requested to handle a JobSubmitted event.

Creating ShuffleMapStage for ShuffleDependency

createShuffleMapStage(
  shuffleDep: ShuffleDependency[_, _, _],
  jobId: Int): ShuffleMapStage

createShuffleMapStage creates a ShuffleMapStage for the given ShuffleDependency as follows:

createShuffleMapStage registers the ShuffleMapStage in the stageIdToStage and shuffleIdToMapStage internal registries.

createShuffleMapStage updateJobIdStageIdMaps.

createShuffleMapStage requests the MapOutputTrackerMaster to check whether it contains the shuffle ID or not.

If not, createShuffleMapStage prints out the following INFO message to the logs and requests the MapOutputTrackerMaster to register the shuffle.

Registering RDD [id] ([creationSite]) as input to shuffle [shuffleId]

DAGScheduler Asks MapOutputTrackerMaster Whether Shuffle Map Output Is Already Tracked

createShuffleMapStage is used when DAGScheduler is requested to find or create a ShuffleMapStage for a given ShuffleDependency.

Cleaning Up After Job and Independent Stages

cleanupStateForJobAndIndependentStages(
  job: ActiveJob): Unit

cleanupStateForJobAndIndependentStages cleans up the state for job and any stages that are not part of any other job.

cleanupStateForJobAndIndependentStages looks the job up in the internal <> registry.

If no stages are found, the following ERROR is printed out to the logs:

No stages registered for job [jobId]

Oterwise, cleanupStateForJobAndIndependentStages uses <> registry to find the stages (the real objects not ids!).

For each stage, cleanupStateForJobAndIndependentStages reads the jobs the stage belongs to.

If the job does not belong to the jobs of the stage, the following ERROR is printed out to the logs:

Job [jobId] not registered for stage [stageId] even though that stage was registered for the job

If the job was the only job for the stage, the stage (and the stage id) gets cleaned up from the registries, i.e. <>, <>, <>, <> and <>.

While removing from <>, you should see the following DEBUG message in the logs:

Removing running stage [stageId]

While removing from <>, you should see the following DEBUG message in the logs:

Removing stage [stageId] from waiting set.

While removing from <>, you should see the following DEBUG message in the logs:

Removing stage [stageId] from failed set.

After all cleaning (using <> as the source registry), if the stage belonged to the one and only job, you should see the following DEBUG message in the logs:

After removal of stage [stageId], remaining stages = [stageIdToStage.size]

The job is removed from <>, <>, <> registries.

The final stage of the job is removed, i.e. ResultStage or ShuffleMapStage.

cleanupStateForJobAndIndependentStages is used in handleTaskCompletion when a ResultTask has completed successfully, failJobAndIndependentStages and markMapStageJobAsFinished.

Marking ShuffleMapStage Job Finished

markMapStageJobAsFinished(
  job: ActiveJob,
  stats: MapOutputStatistics): Unit

markMapStageJobAsFinished marks the active job finished and notifies Spark listeners.

Internally, markMapStageJobAsFinished marks the zeroth partition finished and increases the number of tasks finished in job.

The job listener is notified about the 0th task succeeded.

The <job and independent stages are cleaned up>>.

Ultimately, SparkListenerJobEnd is posted to LiveListenerBus (as <>) for the job, the current time (in millis) and JobSucceeded job result.

markMapStageJobAsFinished is used in handleMapStageSubmitted and handleTaskCompletion.

Finding Or Creating Missing Direct Parent ShuffleMapStages (For ShuffleDependencies) of RDD

getOrCreateParentStages(
  rdd: RDD[_],
  firstJobId: Int): List[Stage]

getOrCreateParentStages <ShuffleDependencies>> of the input rdd and then <ShuffleMapStage stages>> for each ShuffleDependency.

getOrCreateParentStages is used when DAGScheduler is requested to create a ShuffleMapStage or a ResultStage.

Marking Stage Finished

markStageAsFinished(
  stage: Stage,
  errorMessage: Option[String] = None,
  willRetry: Boolean = false): Unit

markStageAsFinished...FIXME

markStageAsFinished is used when...FIXME

Finding or Creating ShuffleMapStage for ShuffleDependency

getOrCreateShuffleMapStage(
  shuffleDep: ShuffleDependency[_, _, _],
  firstJobId: Int): ShuffleMapStage

getOrCreateShuffleMapStage finds a ShuffleMapStage by the shuffleId of the given ShuffleDependency in the shuffleIdToMapStage internal registry and returns it if available.

If not found, getOrCreateShuffleMapStage finds all the missing ancestor shuffle dependencies and creates the missing ShuffleMapStage stages (including one for the input ShuffleDependency).

getOrCreateShuffleMapStage is used when DAGScheduler is requested to find or create missing direct parent ShuffleMapStages of an RDD, find missing parent ShuffleMapStages for a stage, handle a MapStageSubmitted event, and check out stage dependency on a stage.

Finding Missing ShuffleDependencies For RDD

getMissingAncestorShuffleDependencies(
   rdd: RDD[_]): Stack[ShuffleDependency[_, _, _]]

getMissingAncestorShuffleDependencies finds all missing shuffle dependencies for the given RDD traversing its rdd/spark-rdd-lineage.md[RDD lineage].

NOTE: A missing shuffle dependency of a RDD is a dependency not registered in <shuffleIdToMapStage internal registry>>.

Internally, getMissingAncestorShuffleDependencies <> of the input RDD and collects the ones that are not registered in <shuffleIdToMapStage internal registry>>. It repeats the process for the RDDs of the parent shuffle dependencies.

getMissingAncestorShuffleDependencies is used when DAGScheduler is requested to find all ShuffleMapStage stages for a ShuffleDependency.

Finding Direct Parent Shuffle Dependencies of RDD

getShuffleDependencies(
   rdd: RDD[_]): HashSet[ShuffleDependency[_, _, _]]

getShuffleDependencies finds direct parent shuffle dependencies for the given RDD.

getShuffleDependencies Finds Direct Parent ShuffleDependencies (shuffle1 and shuffle2)

Internally, getShuffleDependencies takes the direct rdd/index.md#dependencies[shuffle dependencies of the input RDD] and direct shuffle dependencies of all the parent non-ShuffleDependencies in the dependency chain (aka RDD lineage).

getShuffleDependencies is used when DAGScheduler is requested to find or create missing direct parent ShuffleMapStages (for ShuffleDependencies of a RDD) and find all missing shuffle dependencies for a given RDD.

Failing Job and Independent Single-Job Stages

failJobAndIndependentStages(
  job: ActiveJob,
  failureReason: String,
  exception: Option[Throwable] = None): Unit

failJobAndIndependentStages fails the input job and all the stages that are only used by the job.

Internally, failJobAndIndependentStages uses <jobIdToStageIds internal registry>> to look up the stages registered for the job.

If no stages could be found, you should see the following ERROR message in the logs:

No stages registered for job [id]

Otherwise, for every stage, failJobAndIndependentStages finds the job ids the stage belongs to.

If no stages could be found or the job is not referenced by the stages, you should see the following ERROR message in the logs:

Job [id] not registered for stage [id] even though that stage was registered for the job

Only when there is exactly one job registered for the stage and the stage is in RUNNING state (in runningStages internal registry), TaskScheduler.md#contract[TaskScheduler is requested to cancel the stage's tasks] and <>.

NOTE: failJobAndIndependentStages uses <>, <>, and <> internal registries.

failJobAndIndependentStages is used when...FIXME

Aborting Stage

abortStage(
  failedStage: Stage,
  reason: String,
  exception: Option[Throwable]): Unit

abortStage is an internal method that finds all the active jobs that depend on the failedStage stage and fails them.

Internally, abortStage looks the failedStage stage up in the internal <> registry and exits if there the stage was not registered earlier.

If it was, abortStage finds all the active jobs (in the internal <> registry) with the <failedStage stage>>.

At this time, the completionTime property (of the failed stage's spark-scheduler-StageInfo.md[StageInfo]) is assigned to the current time (millis).

All the active jobs that depend on the failed stage (as calculated above) and the stages that do not belong to other jobs (aka independent stages) are <> (with the failure reason being "Job aborted due to stage failure: [reason]" and the input exception).

If there are no jobs depending on the failed stage, you should see the following INFO message in the logs:

Ignoring failure of [failedStage] because all jobs depending on it are done

abortStage is used when DAGScheduler is requested to handle a TaskSetFailed event, submit a stage, submit missing tasks of a stage, handle a TaskCompletion event.

Checking Out Stage Dependency on Given Stage

stageDependsOn(
  stage: Stage,
  target: Stage): Boolean

stageDependsOn compares two stages and returns whether the stage depends on target stage (i.e. true) or not (i.e. false).

NOTE: A stage A depends on stage B if B is among the ancestors of A.

Internally, stageDependsOn walks through the graph of RDDs of the input stage. For every RDD in the RDD's dependencies (using RDD.dependencies) stageDependsOn adds the RDD of a NarrowDependency to a stack of RDDs to visit while for a ShuffleDependency it <ShuffleMapStage stages for a ShuffleDependency>> for the dependency and the stage's first job id that it later adds to a stack of RDDs to visit if the map stage is ready, i.e. all the partitions have shuffle outputs.

After all the RDDs of the input stage are visited, stageDependsOn checks if the target's RDD is among the RDDs of the stage, i.e. whether the stage depends on target stage.

stageDependsOn is used when DAGScheduler is requested to abort a stage.

Submitting Waiting Child Stages for Execution

submitWaitingChildStages(
  parent: Stage): Unit

submitWaitingChildStages submits for execution all waiting stages for which the input parent Stage.md[Stage] is the direct parent.

NOTE: Waiting stages are the stages registered in <waitingStages internal registry>>.

When executed, you should see the following TRACE messages in the logs:

Checking if any dependencies of [parent] are now runnable
running: [runningStages]
waiting: [waitingStages]
failed: [failedStages]

submitWaitingChildStages finds child stages of the input parent stage, removes them from waitingStages internal registry, and <> one by one sorted by their job ids.

submitWaitingChildStages is used when DAGScheduler is requested to submits missing tasks for a stage and handles a successful ShuffleMapTask completion.

Submitting Stage (with Missing Parents) for Execution

submitStage(
  stage: Stage): Unit

submitStage submits the input stage or its missing parents (if there any stages not computed yet before the input stage could).

NOTE: submitStage is also used to DAGSchedulerEventProcessLoop.md#resubmitFailedStages[resubmit failed stages].

submitStage recursively submits any missing parents of the stage.

Internally, submitStage first finds the earliest-created job id that needs the stage.

NOTE: A stage itself tracks the jobs (their ids) it belongs to (using the internal jobIds registry).

The following steps depend on whether there is a job or not.

If there are no jobs that require the stage, submitStage <> with the reason:

No active job for stage [id]

If however there is a job for the stage, you should see the following DEBUG message in the logs:

submitStage([stage])

submitStage checks the status of the stage and continues when it was not recorded in <>, <> or <> internal registries. It simply exits otherwise.

With the stage ready for submission, submitStage calculates the <stage>> (sorted by their job ids). You should see the following DEBUG message in the logs:

missing: [missing]

When the stage has no parent stages missing, you should see the following INFO message in the logs:

Submitting [stage] ([stage.rdd]), which has no missing parents

submitStage <stage>> (with the earliest-created job id) and finishes.

If however there are missing parent stages for the stage, submitStage <>, and the stage is recorded in the internal <> registry.

submitStage is used recursively for missing parents of the given stage and when DAGScheduler is requested for the following:

  • <> (ResubmitFailedStages event)

  • <> (CompletionEvent event)

  • Handle <>, <> and <> events

Stage Attempts

A single stage can be re-executed in multiple attempts due to fault recovery. The number of attempts is configured (FIXME).

If TaskScheduler reports that a task failed because a map output file from a previous stage was lost, the DAGScheduler resubmits the lost stage. This is detected through a DAGSchedulerEventProcessLoop.md#handleTaskCompletion-FetchFailed[CompletionEvent with FetchFailed], or an <> event. DAGScheduler will wait a small amount of time to see whether other nodes or tasks fail, then resubmit TaskSets for any lost stage(s) that compute the missing tasks.

Please note that tasks from the old attempts of a stage could still be running.

A stage object tracks multiple spark-scheduler-StageInfo.md[StageInfo] objects to pass to Spark listeners or the web UI.

The latest StageInfo for the most recent attempt for a stage is accessible through latestInfo.

Preferred Locations

DAGScheduler computes where to run each task in a stage based on the rdd/index.md#getPreferredLocations[preferred locations of its underlying RDDs], or <>.

Adaptive Query Planning / Adaptive Scheduling

See SPARK-9850 Adaptive execution in Spark for the design document. The work is currently in progress.

DAGScheduler.submitMapStage method is used for adaptive query planning, to run map stages and look at statistics about their outputs before submitting downstream stages.

ScheduledExecutorService daemon services

DAGScheduler uses the following ScheduledThreadPoolExecutors (with the policy of removing cancelled tasks from a work queue at time of cancellation):

  • dag-scheduler-message - a daemon thread pool using j.u.c.ScheduledThreadPoolExecutor with core pool size 1. It is used to post a DAGSchedulerEventProcessLoop.md#ResubmitFailedStages[ResubmitFailedStages] event when DAGSchedulerEventProcessLoop.md#handleTaskCompletion-FetchFailed[FetchFailed is reported].

They are created using ThreadUtils.newDaemonSingleThreadScheduledExecutor method that uses Guava DSL to instantiate a ThreadFactory.

Finding Missing Parent ShuffleMapStages For Stage

getMissingParentStages(
  stage: Stage): List[Stage]

getMissingParentStages finds missing parent ShuffleMapStages in the dependency graph of the input stage (using the breadth-first search algorithm).

Internally, getMissingParentStages starts with the stage's RDD and walks up the tree of all parent RDDs to find <>.

NOTE: A Stage tracks the associated RDD using Stage.md#rdd[rdd property].

NOTE: An uncached partition of a RDD is a partition that has Nil in the <> (which results in no RDD blocks in any of the active storage:BlockManager.md[BlockManager]s on executors).

getMissingParentStages traverses the rdd/index.md#dependencies[parent dependencies of the RDD] and acts according to their type, i.e. ShuffleDependency or NarrowDependency.

NOTE: ShuffleDependency and NarrowDependency are the main top-level Dependencies.

For each NarrowDependency, getMissingParentStages simply marks the corresponding RDD to visit and moves on to a next dependency of a RDD or works on another unvisited parent RDD.

NOTE: NarrowDependency is a RDD dependency that allows for pipelined execution.

getMissingParentStages focuses on ShuffleDependency dependencies.

NOTE: ShuffleDependency is a RDD dependency that represents a dependency on the output of a ShuffleMapStage, i.e. shuffle map stage.

For each ShuffleDependency, getMissingParentStages <ShuffleMapStage stages>>. If the ShuffleMapStage is not available, it is added to the set of missing (map) stages.

NOTE: A ShuffleMapStage is available when all its partitions are computed, i.e. results are available (as blocks).

CAUTION: FIXME...IMAGE with ShuffleDependencies queried

getMissingParentStages is used when DAGScheduler is requested to submit a stage and handle JobSubmitted and MapStageSubmitted events.

Submitting Missing Tasks of Stage

submitMissingTasks(
  stage: Stage,
  jobId: Int): Unit

submitMissingTasks prints out the following DEBUG message to the logs:

submitMissingTasks([stage])

submitMissingTasks requests the given Stage for the missing partitions (partitions that need to be computed).

submitMissingTasks adds the stage to the runningStages internal registry.

submitMissingTasks notifies the OutputCommitCoordinator that stage execution started.

submitMissingTasks determines preferred locations (task locality preferences) of the missing partitions.

submitMissingTasks requests the stage for a new stage attempt.

submitMissingTasks requests the LiveListenerBus to post a SparkListenerStageSubmitted event.

submitMissingTasks uses the closure Serializer to serialize the stage and create a so-called task binary. submitMissingTasks serializes the RDD (of the stage) and either the ShuffleDependency or the compute function based on the type of the stage (ShuffleMapStage or ResultStage, respectively).

submitMissingTasks creates a broadcast variable for the task binary.

Note

That shows how important Broadcasts are for Spark itself to distribute data among executors in a Spark application in the most efficient way.

submitMissingTasks creates tasks for every missing partition:

If there are tasks to submit for execution (i.e. there are missing partitions in the stage), submitMissingTasks prints out the following INFO message to the logs:

Submitting [size] missing tasks from [stage] ([rdd]) (first 15 tasks are for partitions [partitionIds])

submitMissingTasks requests the <> to TaskScheduler.md#submitTasks[submit the tasks for execution] (as a new TaskSet.md[TaskSet]).

With no tasks to submit for execution, submitMissingTasks <>.

submitMissingTasks prints out the following DEBUG messages based on the type of the stage:

Stage [stage] is actually done; (available: [isAvailable],available outputs: [numAvailableOutputs],partitions: [numPartitions])

or

Stage [stage] is actually done; (partitions: [numPartitions])

for ShuffleMapStage and ResultStage, respectively.

In the end, with no tasks to submit for execution, submitMissingTasks <> and exits.

submitMissingTasks is used when DAGScheduler is requested to submit a stage for execution.

Finding Preferred Locations for Missing Partitions

getPreferredLocs(
   rdd: RDD[_],
  partition: Int): Seq[TaskLocation]

getPreferredLocs is simply an alias for the internal (recursive) <>.

getPreferredLocs is used when...FIXME

Finding BlockManagers (Executors) for Cached RDD Partitions (aka Block Location Discovery)

getCacheLocs(
   rdd: RDD[_]): IndexedSeq[Seq[TaskLocation]]

getCacheLocs gives TaskLocations (block locations) for the partitions of the input rdd. getCacheLocs caches lookup results in <> internal registry.

NOTE: The size of the collection from getCacheLocs is exactly the number of partitions in rdd RDD.

NOTE: The size of every TaskLocation collection (i.e. every entry in the result of getCacheLocs) is exactly the number of blocks managed using storage:BlockManager.md[BlockManagers] on executors.

Internally, getCacheLocs finds rdd in the <> internal registry (of partition locations per RDD).

If rdd is not in <> internal registry, getCacheLocs branches per its storage:StorageLevel.md[storage level].

For NONE storage level (i.e. no caching), the result is an empty locations (i.e. no location preference).

For other non-NONE storage levels, getCacheLocs storage:BlockManagerMaster.md#getLocations-block-array[requests BlockManagerMaster for block locations] that are then mapped to TaskLocations with the hostname of the owning BlockManager for a block (of a partition) and the executor id.

NOTE: getCacheLocs uses <> that was defined when <>.

getCacheLocs records the computed block locations per partition (as TaskLocation) in <> internal registry.

NOTE: getCacheLocs requests locations from BlockManagerMaster using storage:BlockId.md#RDDBlockId[RDDBlockId] with the RDD id and the partition indices (which implies that the order of the partitions matters to request proper blocks).

NOTE: DAGScheduler uses TaskLocation.md[TaskLocations] (with host and executor) while storage:BlockManagerMaster.md[BlockManagerMaster] uses storage:BlockManagerId.md[] (to track similar information, i.e. block locations).

getCacheLocs is used when DAGScheduler is requested to find missing parent MapStages and getPreferredLocsInternal.

Finding Placement Preferences for RDD Partition (recursively)

getPreferredLocsInternal(
   rdd: RDD[_],
  partition: Int,
  visited: HashSet[(RDD[_], Int)]): Seq[TaskLocation]

getPreferredLocsInternal first <TaskLocations for the partition of the rdd>> (using <> internal cache) and returns them.

Otherwise, if not found, getPreferredLocsInternal rdd/index.md#preferredLocations[requests rdd for the preferred locations of partition] and returns them.

NOTE: Preferred locations of the partitions of a RDD are also called placement preferences or locality preferences.

Otherwise, if not found, getPreferredLocsInternal finds the first parent NarrowDependency and (recursively) finds TaskLocations.

If all the attempts fail to yield any non-empty result, getPreferredLocsInternal returns an empty collection of TaskLocation.md[TaskLocations].

getPreferredLocsInternal is used when DAGScheduler is requested for the preferred locations for missing partitions.

Stopping DAGScheduler

stop(): Unit

stop stops the internal dag-scheduler-message thread pool, dag-scheduler-event-loop, and TaskScheduler.

stop is used when SparkContext is requested to stop.

Updating Accumulators with Partial Values from Completed Tasks

updateAccumulators(
  event: CompletionEvent): Unit

updateAccumulators merges the partial values of accumulators from a completed task into their "source" accumulators on the driver.

NOTE: It is called by <>.

For each spark-accumulators.md#AccumulableInfo[AccumulableInfo] in the CompletionEvent, a partial value from a task is obtained (from AccumulableInfo.update) and added to the driver's accumulator (using Accumulable.++= method).

For named accumulators with the update value being a non-zero value, i.e. not Accumulable.zero:

  • stage.latestInfo.accumulables for the AccumulableInfo.id is set
  • CompletionEvent.taskInfo.accumulables has a new spark-accumulators.md#AccumulableInfo[AccumulableInfo] added.

CAUTION: FIXME Where are Stage.latestInfo.accumulables and CompletionEvent.taskInfo.accumulables used?

updateAccumulators is used when DAGScheduler is requested to handle a task completion.

checkBarrierStageWithNumSlots

checkBarrierStageWithNumSlots(
  rdd: RDD[_]): Unit

checkBarrierStageWithNumSlots...FIXME

checkBarrierStageWithNumSlots is used when DAGScheduler is requested to create <> and <> stages.

Killing Task

killTaskAttempt(
  taskId: Long,
  interruptThread: Boolean,
  reason: String): Boolean

killTaskAttempt requests the TaskScheduler to kill a task.

killTaskAttempt is used when SparkContext is requested to kill a task.

cleanUpAfterSchedulerStop

cleanUpAfterSchedulerStop(): Unit

cleanUpAfterSchedulerStop...FIXME

cleanUpAfterSchedulerStop is used when DAGSchedulerEventProcessLoop is requested to onStop.

removeExecutorAndUnregisterOutputs

removeExecutorAndUnregisterOutputs(
  execId: String,
  fileLost: Boolean,
  hostToUnregisterOutputs: Option[String],
  maybeEpoch: Option[Long] = None): Unit

removeExecutorAndUnregisterOutputs...FIXME

removeExecutorAndUnregisterOutputs is used when DAGScheduler is requested to handle <> (due to a fetch failure) and <> events.

markMapStageJobsAsFinished

markMapStageJobsAsFinished(
  shuffleStage: ShuffleMapStage): Unit

markMapStageJobsAsFinished...FIXME

markMapStageJobsAsFinished is used when DAGScheduler is requested to submit missing tasks (of a ShuffleMapStage that has just been computed) and handle a task completion (of a ShuffleMapStage).

updateJobIdStageIdMaps

updateJobIdStageIdMaps(
  jobId: Int,
  stage: Stage): Unit

updateJobIdStageIdMaps...FIXME

updateJobIdStageIdMaps is used when DAGScheduler is requested to create ShuffleMapStage and ResultStage stages.

executorHeartbeatReceived

executorHeartbeatReceived(
  execId: String,
  // (taskId, stageId, stageAttemptId, accumUpdates)
  accumUpdates: Array[(Long, Int, Int, Seq[AccumulableInfo])],
  blockManagerId: BlockManagerId,
  // (stageId, stageAttemptId) -> metrics
  executorUpdates: mutable.Map[(Int, Int), ExecutorMetrics]): Boolean

executorHeartbeatReceived posts a SparkListenerExecutorMetricsUpdate (to listenerBus) and informs BlockManagerMaster that blockManagerId block manager is alive (by posting BlockManagerHeartbeat).

executorHeartbeatReceived is used when TaskSchedulerImpl is requested to handle an executor heartbeat.

postTaskEnd

postTaskEnd(
  event: CompletionEvent): Unit

postTaskEnd...FIXME

postTaskEnd is used when DAGScheduler is requested to handle a task completion.

Event Handlers

AllJobsCancelled Event Handler

doCancelAllJobs(): Unit

doCancelAllJobs...FIXME

doCancelAllJobs is used when DAGSchedulerEventProcessLoop is requested to handle an AllJobsCancelled event and onError.

BeginEvent Event Handler

handleBeginEvent(
  task: Task[_],
  taskInfo: TaskInfo): Unit

handleBeginEvent...FIXME

handleBeginEvent is used when DAGSchedulerEventProcessLoop is requested to handle a BeginEvent event.

Handling Task Completion Event

handleTaskCompletion(
  event: CompletionEvent): Unit

DAGScheduler and CompletionEvent

handleTaskCompletion handles a CompletionEvent.

handleTaskCompletion starts by scheduler:OutputCommitCoordinator.md#taskCompleted[notifying OutputCommitCoordinator that a task completed].

handleTaskCompletion executor:TaskMetrics.md#fromAccumulators[re-creates TaskMetrics] (using <accumUpdates accumulators of the input event>>).

NOTE: executor:TaskMetrics.md[] can be empty when the task has failed.

handleTaskCompletion announces task completion application-wide (by posting a ROOT:SparkListener.md#SparkListenerTaskEnd[SparkListenerTaskEnd] to scheduler:LiveListenerBus.md[]).

handleTaskCompletion checks the stage of the task out in the scheduler:DAGScheduler.md#stageIdToStage[stageIdToStage internal registry] and if not found, it simply exits.

handleTaskCompletion branches off per TaskEndReason (as event.reason).

.handleTaskCompletion Branches per TaskEndReason [cols="1,2",options="header",width="100%"] |=== | TaskEndReason | Description

| <> | Acts according to the type of the task that completed, i.e. <> and <>.

<>
<>

| ExceptionFailure | scheduler:DAGScheduler.md#updateAccumulators[Updates accumulators] (with partial values from the task).

| ExecutorLostFailure | Does nothing

| TaskCommitDenied | Does nothing

| TaskKilled | Does nothing

| TaskResultLost | Does nothing

| UnknownReason | Does nothing |===

=== [[handleTaskCompletion-Success]] Handling Successful Task Completion

When a task has finished successfully (i.e. Success end reason), handleTaskCompletion marks the partition as no longer pending (i.e. the partition the task worked on is removed from pendingPartitions of the stage).

NOTE: A Stage tracks its own pending partitions using scheduler:Stage.md#pendingPartitions[pendingPartitions property].

handleTaskCompletion branches off given the type of the task that completed, i.e. <> and <>.

==== [[handleTaskCompletion-Success-ResultTask]] Handling Successful ResultTask Completion

For scheduler:ResultTask.md[ResultTask], the stage is assumed a scheduler:ResultStage.md[ResultStage].

handleTaskCompletion finds the ActiveJob associated with the ResultStage.

NOTE: scheduler:ResultStage.md[ResultStage] tracks the optional ActiveJob as scheduler:ResultStage.md#activeJob[activeJob property]. There could only be one active job for a ResultStage.

If there is no job for the ResultStage, you should see the following INFO message in the logs:

Ignoring result from [task] because its job has finished

Otherwise, when the ResultStage has a ActiveJob, handleTaskCompletion checks the status of the partition output for the partition the ResultTask ran for.

NOTE: ActiveJob tracks task completions in finished property with flags for every partition in a stage. When the flag for a partition is enabled (i.e. true), it is assumed that the partition has been computed (and no results from any ResultTask are expected and hence simply ignored).

CAUTION: FIXME Describe why could a partition has more ResultTask running.

handleTaskCompletion ignores the CompletionEvent when the partition has already been marked as completed for the stage and simply exits.

handleTaskCompletion scheduler:DAGScheduler.md#updateAccumulators[updates accumulators].

The partition for the ActiveJob (of the ResultStage) is marked as computed and the number of partitions calculated increased.

NOTE: ActiveJob tracks what partitions have already been computed and their number.

If the ActiveJob has finished (when the number of partitions computed is exactly the number of partitions in a stage) handleTaskCompletion does the following (in order):

  1. scheduler:DAGScheduler.md#markStageAsFinished[Marks ResultStage computed].
  2. scheduler:DAGScheduler.md#cleanupStateForJobAndIndependentStages[Cleans up after ActiveJob and independent stages].
  3. Announces the job completion application-wide (by posting a ROOT:SparkListener.md#SparkListenerJobEnd[SparkListenerJobEnd] to scheduler:LiveListenerBus.md[]).

In the end, handleTaskCompletion notifies JobListener of the ActiveJob that the task succeeded.

NOTE: A task succeeded notification holds the output index and the result.

When the notification throws an exception (because it runs user code), handleTaskCompletion notifies JobListener about the failure (wrapping it inside a SparkDriverExecutionException exception).

==== [[handleTaskCompletion-Success-ShuffleMapTask]] Handling Successful ShuffleMapTask Completion

For scheduler:ShuffleMapTask.md[ShuffleMapTask], the stage is assumed a scheduler:ShuffleMapStage.md[ShuffleMapStage].

handleTaskCompletion scheduler:DAGScheduler.md#updateAccumulators[updates accumulators].

The task's result is assumed scheduler:MapStatus.md[MapStatus] that knows the executor where the task has finished.

You should see the following DEBUG message in the logs:

DEBUG DAGScheduler: ShuffleMapTask finished on [execId]

If the executor is registered in scheduler:DAGScheduler.md#failedEpoch[failedEpoch internal registry] and the epoch of the completed task is not greater than that of the executor (as in failedEpoch registry), you should see the following INFO message in the logs:

INFO DAGScheduler: Ignoring possibly bogus [task] completion from executor [executorId]

Otherwise, handleTaskCompletion scheduler:ShuffleMapStage.md#addOutputLoc[registers the MapStatus result for the partition with the stage] (of the completed task).

handleTaskCompletion does more processing only if the ShuffleMapStage is registered as still running (in scheduler:DAGScheduler.md#runningStages[runningStages internal registry]) and the scheduler:Stage.md#pendingPartitions[ShuffleMapStage stage has no pending partitions to compute].

The ShuffleMapStage is <>.

You should see the following INFO messages in the logs:

INFO DAGScheduler: looking for newly runnable stages
INFO DAGScheduler: running: [runningStages]
INFO DAGScheduler: waiting: [waitingStages]
INFO DAGScheduler: failed: [failedStages]

handleTaskCompletion scheduler:MapOutputTrackerMaster.md#registerMapOutputs[registers the shuffle map outputs of the ShuffleDependency with MapOutputTrackerMaster] (with the epoch incremented) and scheduler:DAGScheduler.md#clearCacheLocs[clears internal cache of the stage's RDD block locations].

NOTE: scheduler:MapOutputTrackerMaster.md[MapOutputTrackerMaster] is given when scheduler:DAGScheduler.md#creating-instance[DAGScheduler is created].

If the scheduler:ShuffleMapStage.md#isAvailable[ShuffleMapStage stage is ready], all scheduler:ShuffleMapStage.md#mapStageJobs[active jobs of the stage] (aka map-stage jobs) are scheduler:DAGScheduler.md#markMapStageJobAsFinished[marked as finished] (with scheduler:MapOutputTrackerMaster.md#getStatistics[MapOutputStatistics from MapOutputTrackerMaster for the ShuffleDependency]).

NOTE: A ShuffleMapStage stage is ready (aka available) when all partitions have shuffle outputs, i.e. when their tasks have completed.

Eventually, handleTaskCompletion scheduler:DAGScheduler.md#submitWaitingChildStages[submits waiting child stages (of the ready ShuffleMapStage)].

If however the ShuffleMapStage is not ready, you should see the following INFO message in the logs:

INFO DAGScheduler: Resubmitting [shuffleStage] ([shuffleStage.name]) because some of its tasks had failed: [missingPartitions]

In the end, handleTaskCompletion scheduler:DAGScheduler.md#submitStage[submits the ShuffleMapStage for execution].

=== [[handleTaskCompletion-Resubmitted]] TaskEndReason: Resubmitted

For Resubmitted case, you should see the following INFO message in the logs:

INFO Resubmitted [task], so marking it as still running

The task (by task.partitionId) is added to the collection of pending partitions of the stage (using stage.pendingPartitions).

TIP: A stage knows how many partitions are yet to be calculated. A task knows about the partition id for which it was launched.

=== [[handleTaskCompletion-FetchFailed]] Task Failed with FetchFailed Exception -- TaskEndReason: FetchFailed

[source, scala]

FetchFailed( bmAddress: BlockManagerId, shuffleId: Int, mapId: Int, reduceId: Int, message: String) extends TaskFailedReason


.FetchFailed Properties [cols="1,2",options="header",width="100%"] |=== | Name | Description

| bmAddress | storage:BlockManagerId.md[]

| shuffleId | Used when...

| mapId | Used when...

| reduceId | Used when...

| failureMessage | Used when... |===

NOTE: A task knows about the id of the stage it belongs to.

When FetchFailed happens, stageIdToStage is used to access the failed stage (using task.stageId and the task is available in event in handleTaskCompletion(event: CompletionEvent)). shuffleToMapStage is used to access the map stage (using shuffleId).

If failedStage.latestInfo.attemptId != task.stageAttemptId, you should see the following INFO in the logs:

INFO Ignoring fetch failure from [task] as it's from [failedStage] attempt [task.stageAttemptId] and there is a more recent attempt for that stage (attempt ID [failedStage.latestInfo.attemptId]) running

CAUTION: FIXME What does failedStage.latestInfo.attemptId != task.stageAttemptId mean?

And the case finishes. Otherwise, the case continues.

If the failed stage is in runningStages, the following INFO message shows in the logs:

INFO Marking [failedStage] ([failedStage.name]) as failed due to a fetch failure from [mapStage] ([mapStage.name])

markStageAsFinished(failedStage, Some(failureMessage)) is called.

CAUTION: FIXME What does markStageAsFinished do?

If the failed stage is not in runningStages, the following DEBUG message shows in the logs:

DEBUG Received fetch failure from [task], but its from [failedStage] which is no longer running

When disallowStageRetryForTest is set, abortStage(failedStage, "Fetch failure will not retry stage due to testing config", None) is called.

CAUTION: FIXME Describe disallowStageRetryForTest and abortStage.

If the scheduler:Stage.md#failedOnFetchAndShouldAbort[number of fetch failed attempts for the stage exceeds the allowed number], the scheduler:DAGScheduler.md#abortStage[failed stage is aborted] with the reason:

[failedStage] ([name]) has failed the maximum allowable number of times: 4. Most recent failure reason: [failureMessage]

If there are no failed stages reported (scheduler:DAGScheduler.md#failedStages[DAGScheduler.failedStages] is empty), the following INFO shows in the logs:

INFO Resubmitting [mapStage] ([mapStage.name]) and [failedStage] ([failedStage.name]) due to fetch failure

And the following code is executed:

messageScheduler.schedule(
  new Runnable {
    override def run(): Unit = eventProcessLoop.post(ResubmitFailedStages)
  }, DAGScheduler.RESUBMIT_TIMEOUT, TimeUnit.MILLISECONDS)

CAUTION: FIXME What does the above code do?

For all the cases, the failed stage and map stages are both added to the internal scheduler:DAGScheduler.md#failedStages[registry of failed stages].

If mapId (in the FetchFailed object for the case) is provided, the map stage output is cleaned up (as it is broken) using mapStage.removeOutputLoc(mapId, bmAddress) and scheduler:MapOutputTracker.md#unregisterMapOutput[MapOutputTrackerMaster.unregisterMapOutput(shuffleId, mapId, bmAddress)] methods.

CAUTION: FIXME What does mapStage.removeOutputLoc do?

If BlockManagerId (as bmAddress in the FetchFailed object) is defined, handleTaskCompletion <> (with filesLost enabled and maybeEpoch from the scheduler:Task.md#epoch[Task] that completed).

handleTaskCompletion is used when DAGSchedulerEventProcessLoop is requested to handle a CompletionEvent event.

ExecutorAdded Event Handler

handleExecutorAdded(
  execId: String,
  host: String): Unit

handleExecutorAdded...FIXME

handleExecutorAdded is used when DAGSchedulerEventProcessLoop is requested to handle an ExecutorAdded event.

ExecutorLost Event Handler

handleExecutorLost(
  execId: String,
  workerLost: Boolean): Unit

handleExecutorLost checks whether the input optional maybeEpoch is defined and if not requests the scheduler:MapOutputTracker.md#getEpoch[current epoch from MapOutputTrackerMaster].

NOTE: MapOutputTrackerMaster is passed in (as mapOutputTracker) when scheduler:DAGScheduler.md#creating-instance[DAGScheduler is created].

CAUTION: FIXME When is maybeEpoch passed in?

.DAGScheduler.handleExecutorLost image::dagscheduler-handleExecutorLost.png[align="center"]

Recurring ExecutorLost events lead to the following repeating DEBUG message in the logs:

DEBUG Additional executor lost message for [execId] (epoch [currentEpoch])

NOTE: handleExecutorLost handler uses DAGScheduler's failedEpoch and FIXME internal registries.

Otherwise, when the executor execId is not in the scheduler:DAGScheduler.md#failedEpoch[list of executor lost] or the executor failure's epoch is smaller than the input maybeEpoch, the executor's lost event is recorded in scheduler:DAGScheduler.md#failedEpoch[failedEpoch internal registry].

CAUTION: FIXME Describe the case above in simpler non-technical words. Perhaps change the order, too.

You should see the following INFO message in the logs:

INFO Executor lost: [execId] (epoch [epoch])

storage:BlockManagerMaster.md#removeExecutor[BlockManagerMaster is requested to remove the lost executor execId].

CAUTION: FIXME Review what's filesLost.

handleExecutorLost exits unless the ExecutorLost event was for a map output fetch operation (and the input filesLost is true) or deploy:ExternalShuffleService.md[external shuffle service] is not used.

In such a case, you should see the following INFO message in the logs:

INFO Shuffle files lost for executor: [execId] (epoch [epoch])

handleExecutorLost walks over all scheduler:ShuffleMapStage.md[ShuffleMapStage]s in scheduler:DAGScheduler.md#shuffleToMapStage[DAGScheduler's shuffleToMapStage internal registry] and do the following (in order):

  1. ShuffleMapStage.removeOutputsOnExecutor(execId) is called
  2. scheduler:MapOutputTrackerMaster.md#registerMapOutputs[MapOutputTrackerMaster.registerMapOutputs(shuffleId, stage.outputLocInMapOutputTrackerFormat(), changeEpoch = true)] is called.

In case scheduler:DAGScheduler.md#shuffleToMapStage[DAGScheduler's shuffleToMapStage internal registry] has no shuffles registered, scheduler:MapOutputTrackerMaster.md#incrementEpoch[MapOutputTrackerMaster is requested to increment epoch].

Ultimatelly, DAGScheduler scheduler:DAGScheduler.md#clearCacheLocs[clears the internal cache of RDD partition locations].

handleExecutorLost is used when DAGSchedulerEventProcessLoop is requested to handle an ExecutorLost event.

GettingResultEvent Event Handler

handleGetTaskResult(
  taskInfo: TaskInfo): Unit

handleGetTaskResult...FIXME

handleGetTaskResult is used when DAGSchedulerEventProcessLoop is requested to handle a GettingResultEvent event.

JobCancelled Event Handler

handleJobCancellation(
  jobId: Int,
  reason: Option[String]): Unit

handleJobCancellation looks up the active job for the input job ID (in jobIdToActiveJob internal registry) and fails it and all associated independent stages with failure reason:

Job [jobId] cancelled [reason]

When the input job ID is not found, handleJobCancellation prints out the following DEBUG message to the logs:

Trying to cancel unregistered job [jobId]

handleJobCancellation is used when DAGScheduler is requested to handle a JobCancelled event, doCancelAllJobs, handleJobGroupCancelled, handleStageCancellation.

JobGroupCancelled Event Handler

handleJobGroupCancelled(
  groupId: String): Unit

handleJobGroupCancelled finds active jobs in a group and cancels them.

Internally, handleJobGroupCancelled computes all the active jobs (registered in the internal collection of active jobs) that have spark.jobGroup.id scheduling property set to groupId.

handleJobGroupCancelled then cancels every active job in the group one by one and the cancellation reason:

part of cancelled job group [groupId]

handleJobGroupCancelled is used when DAGScheduler is requested to handle JobGroupCancelled event.

Handling JobSubmitted Event

handleJobSubmitted(
  jobId: Int,
  finalRDD: RDD[_],
  func: (TaskContext, Iterator[_]) => _,
  partitions: Array[Int],
  callSite: CallSite,
  listener: JobListener,
  properties: Properties): Unit

handleJobSubmitted creates a ResultStage (as finalStage in the picture below) for the given RDD, func, partitions, jobId and callSite.

DAGScheduler.handleJobSubmitted Method

handleJobSubmitted creates an ActiveJob for the ResultStage.

handleJobSubmitted clears the internal cache of RDD partition locations.

Important

FIXME Why is this clearing here so important?

handleJobSubmitted prints out the following INFO messages to the logs (with missingParentStages):

Got job [id] ([callSite]) with [number] output partitions
Final stage: [stage] ([name])
Parents of final stage: [parents]
Missing parents: [missingParentStages]

handleJobSubmitted registers the new ActiveJob in jobIdToActiveJob and activeJobs internal registries.

handleJobSubmitted requests the ResultStage to associate itself with the ActiveJob.

handleJobSubmitted uses the jobIdToStageIds internal registry to find all registered stages for the given jobId. handleJobSubmitted uses the stageIdToStage internal registry to request the Stages for the latestInfo.

In the end, handleJobSubmitted posts a SparkListenerJobStart message to the LiveListenerBus and submits the ResultStage.

handleJobSubmitted is used when DAGSchedulerEventProcessLoop is requested to handle a JobSubmitted event.

MapStageSubmitted Event Handler

handleMapStageSubmitted(
  jobId: Int,
  dependency: ShuffleDependency[_, _, _],
  callSite: CallSite,
  listener: JobListener,
  properties: Properties): Unit

MapStageSubmitted Event Handling

Note

MapStageSubmitted event processing is very similar to <> events.

handleMapStageSubmitted finds or creates a new ShuffleMapStage for the input ShuffleDependency and jobId.

handleMapStageSubmitted creates an ActiveJob.

handleMapStageSubmitted clears the internal cache of RDD partition locations.

Important

FIXME Why is this clearing here so important?

handleMapStageSubmitted prints out the following INFO messages to the logs:

Got map stage job [id] ([callSite]) with [number] output partitions
Final stage: [stage] ([name])
Parents of final stage: [parents]
Missing parents: [missingParentStages]

handleMapStageSubmitted registers the new job in jobIdToActiveJob and activeJobs internal registries, and with the final ShuffleMapStage.

Note

ShuffleMapStage can have multiple ActiveJobs registered.

handleMapStageSubmitted finds all the registered stages for the input jobId and collects their latest StageInfo.

In the end, handleMapStageSubmitted posts SparkListenerJobStart message to LiveListenerBus and submits the ShuffleMapStage.

When the ShuffleMapStage is available already, handleMapStageSubmitted marks the job finished.

When handleMapStageSubmitted could not find or create a ShuffleMapStage, handleMapStageSubmitted prints out the following WARN message to the logs.

Creating new stage failed due to exception - job: [id]

handleMapStageSubmitted notifies listener about the job failure and exits.

handleMapStageSubmitted is used when DAGSchedulerEventProcessLoop is requested to handle a MapStageSubmitted event.

ResubmitFailedStages Event Handler

resubmitFailedStages(): Unit

resubmitFailedStages iterates over the internal collection of failed stages and submits them.

Note

resubmitFailedStages does nothing when there are no failed stages reported.

resubmitFailedStages prints out the following INFO message to the logs:

Resubmitting failed stages

resubmitFailedStages clears the internal cache of RDD partition locations and makes a copy of the collection of failed stages to track failed stages afresh.

Note

At this point DAGScheduler has no failed stages reported.

The previously-reported failed stages are sorted by the corresponding job ids in incremental order and resubmitted.

resubmitFailedStages is used when DAGSchedulerEventProcessLoop is requested to handle a ResubmitFailedStages event.

SpeculativeTaskSubmitted Event Handler

handleSpeculativeTaskSubmitted(): Unit

handleSpeculativeTaskSubmitted...FIXME

handleSpeculativeTaskSubmitted is used when DAGSchedulerEventProcessLoop is requested to handle a SpeculativeTaskSubmitted event.

StageCancelled Event Handler

handleStageCancellation(): Unit

handleStageCancellation...FIXME

handleStageCancellation is used when DAGSchedulerEventProcessLoop is requested to handle a StageCancelled event.

TaskSetFailed Event Handler

handleTaskSetFailed(): Unit

handleTaskSetFailed...FIXME

handleTaskSetFailed is used when DAGSchedulerEventProcessLoop is requested to handle a TaskSetFailed event.

WorkerRemoved Event Handler

handleWorkerRemoved(
  workerId: String,
  host: String,
  message: String): Unit

handleWorkerRemoved...FIXME

handleWorkerRemoved is used when DAGSchedulerEventProcessLoop is requested to handle a WorkerRemoved event.

Internal Properties

failedEpoch

The lookup table of lost executors and the epoch of the event.

failedStages

Stages that failed due to fetch failures (when a DAGSchedulerEventProcessLoop.md#handleTaskCompletion-FetchFailed[task fails with FetchFailed exception]).

jobIdToActiveJob

The lookup table of ActiveJobs per job id.

jobIdToStageIds

The lookup table of all stages per ActiveJob id

nextJobId Counter

nextJobId: AtomicInteger

nextJobId is a Java AtomicInteger for job IDs.

nextJobId starts at 0.

Used when DAGScheduler is requested for numTotalJobs, to submitJob, runApproximateJob and submitMapStage.

nextStageId

The next stage id counting from 0.

Used when DAGScheduler creates a <> and a <>. It is the key in <>.

runningStages

The set of stages that are currently "running".

A stage is added when <> gets executed (without first checking if the stage has not already been added).

shuffleIdToMapStage

A lookup table of ShuffleMapStages by ShuffleDependency

stageIdToStage

A lookup table of stages by stage ID

Used when DAGScheduler creates a shuffle map stage, creates a result stage, cleans up job state and independent stages, is informed that a task is started, a taskset has failed, a job is submitted (to compute a ResultStage), a map stage was submitted, a task has completed or a stage was cancelled, updates accumulators, aborts a stage and fails a job and independent stages.

waitingStages

Stages with parents to be computed

Event Posting Methods

Posting AllJobsCancelled

Posts an AllJobsCancelled

Used when SparkContext is requested to cancel all running or scheduled Spark jobs

Posting JobCancelled

Posts a JobCancelled

Used when SparkContext or JobWaiter are requested to cancel a Spark job

Posting JobGroupCancelled

Posts a JobGroupCancelled

Used when SparkContext is requested to cancel a job group

Posting StageCancelled

Posts a StageCancelled

Used when SparkContext is requested to cancel a stage

Posting ExecutorAdded

Posts an ExecutorAdded

Used when TaskSchedulerImpl is requested to handle resource offers (and a new executor is found in the resource offers)

Posting ExecutorLost

Posts a ExecutorLost

Used when TaskSchedulerImpl is requested to handle a task status update (and a task gets lost which is used to indicate that the executor got broken and hence should be considered lost) or executorLost

Posting JobSubmitted

Posts a JobSubmitted

Used when SparkContext is requested to run an approximate job

Posting SpeculativeTaskSubmitted

Posts a SpeculativeTaskSubmitted

Used when TaskSetManager is requested to checkAndSubmitSpeculatableTask

Posting MapStageSubmitted

Posts a MapStageSubmitted

Used when SparkContext is requested to submit a MapStage for execution

Posting CompletionEvent

Posts a CompletionEvent

Used when TaskSetManager is requested to handleSuccessfulTask, handleFailedTask, and executorLost

Posting GettingResultEvent

Posts a GettingResultEvent

Used when TaskSetManager is requested to handle a task fetching result

Posting TaskSetFailed

Posts a TaskSetFailed

Used when TaskSetManager is requested to abort

Posting BeginEvent

Posts a BeginEvent

Used when TaskSetManager is requested to start a task

Posting WorkerRemoved

Posts a WorkerRemoved

Used when TaskSchedulerImpl is requested to handle a removed worker event

Logging

Enable ALL logging level for org.apache.spark.scheduler.DAGScheduler logger to see what happens inside.

Add the following line to conf/log4j.properties:

log4j.logger.org.apache.spark.scheduler.DAGScheduler=ALL

Refer to Logging.


Last update: 2020-10-12