Inside Creating SparkContext

This document describes what happens when you create a new SparkContext.

import org.apache.spark.{SparkConf, SparkContext}

// 1. Create Spark configuration
val conf = new SparkConf()
  .setAppName("SparkMe Application")
  .setMaster("local[*]")  // local mode

// 2. Create Spark context
val sc = new SparkContext(conf)
The example uses Spark in local mode, but the initialization with the other cluster modes would follow similar steps.

Creating SparkContext instance starts by setting the internal allowMultipleContexts field with the value of spark.driver.allowMultipleContexts and marking this SparkContext instance as partially constructed. It makes sure that no other thread is creating a SparkContext instance in this JVM. It does so by synchronizing on SPARK_CONTEXT_CONSTRUCTOR_LOCK and using the internal atomic reference activeContext (that eventually has a fully-created SparkContext instance).

The entire code of SparkContext that creates a fully-working SparkContext instance is between two statements:

SparkContext.markPartiallyConstructed(this, allowMultipleContexts)

// the SparkContext code goes here

SparkContext.setActiveContext(this, allowMultipleContexts)

startTime is set to the current time in milliseconds.

stopped internal flag is set to false.

The very first information printed out is the version of Spark as an INFO message:

INFO SparkContext: Running Spark version 2.0.0-SNAPSHOT
You can use version method to learn about the current Spark version or org.apache.spark.SPARK_VERSION value.

The current user name is computed.

FIXME Where is sparkUser used?

It saves the input SparkConf (as _conf).

FIXME Review _conf.validateSettings()

It ensures that the first mandatory setting - spark.master is defined. SparkException is thrown if not.

A master URL must be set in your configuration

It ensures that the other mandatory setting - is defined. SparkException is thrown if not.

An application name must be set in your configuration

For Spark on YARN in cluster deploy mode, it checks existence of SparkException is thrown if it does not exist.

Detected yarn cluster mode, but isn't running on a cluster. Deployment to YARN is not supported directly by SparkContext. Please use spark-submit.
FIXME How to "trigger" the exception? What are the steps?

When spark.logConf is enabled SparkConf.toDebugString is called.

SparkConf.toDebugString is called very early in the initialization process and other settings configured afterwards are not included. Use sc.getConf.toDebugString once SparkContext is initialized.

The driver’s host and port are set if missing. becomes the value of Utils.localHostName (or an exception is thrown) while spark.driver.port is set to 0. and spark.driver.port are expected to be set on the driver. It is later asserted by SparkEnv. setting is set to driver.

Use sc.getConf.get("") to know where the code is executed — driver or executors.

It sets the jars and files based on spark.jars and spark.files, respectively. These are files that are required for proper task execution on executors.

If event logging is enabled, i.e. spark.eventLog.enabled flag is true, the internal field _eventLogDir is set to the value of spark.eventLog.dir setting or the default value /tmp/spark-events.

Also, if spark.eventLog.compress is enabled (it is not by default), the short name of the CompressionCodec is assigned to _eventLogCodec. The config key is (default: lz4).

Read about compression codecs in Compression.

Creating LiveListenerBus

SparkContext creates a LiveListenerBus.

Creating Live AppStatusStore

SparkContext requests AppStatusStore to create a live store (i.e. the AppStatusStore for a live Spark application) and requests LiveListenerBus to add the AppStatusListener to the status queue.

The current AppStatusStore is available as statusStore property of the SparkContext.

Creating SparkEnv

SparkContext creates a SparkEnv and requests SparkEnv to use the instance as the default SparkEnv.

FIXME Describe the following steps.

MetadataCleaner is created.

FIXME What’s MetadataCleaner?

Creating SparkStatusTracker

SparkContext creates a SparkStatusTracker (with itself and the AppStatusStore).

Creating ConsoleProgressBar

SparkContext creates the optional ConsoleProgressBar when spark.ui.showConsoleProgress property is enabled and the INFO logging level for SparkContext is disabled.

Creating SparkUI

SparkContext creates a SparkUI when spark.ui.enabled configuration property is enabled (i.e. true) with the following:

spark.ui.enabled Spark property is assumed enabled when undefined.
FIXME Where’s _ui used?

A Hadoop configuration is created. See Hadoop Configuration.

If there are jars given through the SparkContext constructor, they are added using addJar.

If there were files specified, they are added using addFile.

At this point in time, the amount of memory to allocate to each executor (as _executorMemory) is calculated. It is the value of spark.executor.memory setting, or SPARK_EXECUTOR_MEMORY environment variable (or currently-deprecated SPARK_MEM), or defaults to 1024.

_executorMemory is later available as sc.executorMemory and used for LOCAL_CLUSTER_REGEX, Spark Standalone’s SparkDeploySchedulerBackend, to set executorEnvs("SPARK_EXECUTOR_MEMORY"), MesosSchedulerBackend, CoarseMesosSchedulerBackend.

The value of SPARK_PREPEND_CLASSES environment variable is included in executorEnvs.


  • What’s _executorMemory?

  • What’s the unit of the value of _executorMemory exactly?

  • What are "SPARK_TESTING", "spark.testing"? How do they contribute to executorEnvs?

  • What’s executorEnvs?

The Mesos scheduler backend’s configuration is included in executorEnvs, i.e. SPARK_EXECUTOR_MEMORY, _conf.getExecutorEnv, and SPARK_USER.

SparkContext registers HeartbeatReceiver RPC endpoint.

SparkContext object is requested to create the SchedulerBackend with the TaskScheduler (for the given master URL) and the result becomes the internal _schedulerBackend and _taskScheduler.

The internal _schedulerBackend and _taskScheduler are used by schedulerBackend and taskScheduler methods, respectively.

DAGScheduler is created (as _dagScheduler).

SparkContext sends a blocking TaskSchedulerIsSet message to HeartbeatReceiver RPC endpoint (to inform that the TaskScheduler is now available).

Starting TaskScheduler

SparkContext starts TaskScheduler.

Setting Spark Application’s and Execution Attempt’s IDs — _applicationId and _applicationAttemptId

SparkContext sets the internal fields — _applicationId and _applicationAttemptId — (using applicationId and applicationAttemptId methods from the TaskScheduler Contract).

SparkContext requests TaskScheduler for the unique identifier of a Spark application (that is currently only implemented by TaskSchedulerImpl that uses SchedulerBackend to request the identifier).
The unique identifier of a Spark application is used to initialize SparkUI and BlockManager.
_applicationAttemptId is used when SparkContext is requested for the unique identifier of execution attempt of a Spark application and when EventLoggingListener is created.

Setting Spark Property in SparkConf

SparkContext sets property to be the unique identifier of a Spark application and, if enabled, passes it on to SparkUI.

Initializing BlockManager

Starting MetricsSystem

SparkContext requests the MetricsSystem to start.

SparkContext starts MetricsSystem after setting Spark property as MetricsSystem uses it to build unique identifiers fo metrics sources.

Requesting JSON Servlet Handler

SparkContext requests the MetricsSystem for a JSON servlet handler and requests the SparkUI to attach it.

_eventLogger is created and started if isEventLogEnabled. It uses EventLoggingListener that gets registered to LiveListenerBus.

FIXME Why is _eventLogger required to be the internal field of SparkContext? Where is this used?

If dynamic allocation is enabled, ExecutorAllocationManager is created (as _executorAllocationManager) and immediately started.

_executorAllocationManager is exposed (as a method) to YARN scheduler backends to reset their state to the initial state.

If spark.cleaner.referenceTracking Spark property is enabled (i.e. true), SparkContext creates ContextCleaner (as _cleaner) and started immediately. Otherwise, _cleaner is empty.

spark.cleaner.referenceTracking Spark property is enabled by default.
FIXME It’d be quite useful to have all the properties with their default values in sc.getConf.toDebugString, so when a configuration is not included but does change Spark runtime configuration, it should be added to _conf.

postEnvironmentUpdate is called that posts SparkListenerEnvironmentUpdate message on LiveListenerBus with information about Task Scheduler’s scheduling mode, added jar and file paths, and other environmental details. They are displayed in web UI’s Environment tab.

SparkListenerApplicationStart message is posted to LiveListenerBus (using the internal postApplicationStart method).

TaskScheduler.postStartHook does nothing by default, but custom implementations offer more advanced features, i.e. TaskSchedulerImpl blocks the current thread until SchedulerBackend is ready. There is also YarnClusterScheduler for Spark on YARN in cluster deploy mode.

Registering Metrics Sources

SparkContext requests MetricsSystem to register metrics sources for the following services:

Adding Shutdown Hook

SparkContext adds a shutdown hook (using ShutdownHookManager.addShutdownHook()).

You should see the following DEBUG message in the logs:

DEBUG Adding shutdown hook
FIXME ShutdownHookManager.addShutdownHook()

Any non-fatal Exception leads to termination of the Spark context instance.

FIXME What does NonFatal represent in Scala?
FIXME Finish me

Initializing nextShuffleId and nextRddId Internal Counters

nextShuffleId and nextRddId start with 0.

FIXME Where are nextShuffleId and nextRddId used?

A new instance of Spark context is created and ready for operation.

Loading External Cluster Manager for URL (getClusterManager method)

getClusterManager(url: String): Option[ExternalClusterManager]

getClusterManager loads ExternalClusterManager that can handle the input url.

If there are two or more external cluster managers that could handle url, a SparkException is thrown:

Multiple Cluster Managers ([serviceLoaders]) registered for the url [url].
getClusterManager uses Java’s ServiceLoader.load method.
getClusterManager is used to find a cluster manager for a master URL when creating a SchedulerBackend and a TaskScheduler for the driver.


setupAndStartListenerBus(): Unit

setupAndStartListenerBus is an internal method that reads spark.extraListeners setting from the current SparkConf to create and register SparkListenerInterface listeners.

It expects that the class name represents a SparkListenerInterface listener with one of the following constructors (in this order):

  • a single-argument constructor that accepts SparkConf

  • a zero-argument constructor

setupAndStartListenerBus registers every listener class.

You should see the following INFO message in the logs:

INFO Registered listener [className]

It starts LiveListenerBus and records it in the internal _listenerBusStarted.

When no single-SparkConf or zero-argument constructor could be found for a class name in spark.extraListeners setting, a SparkException is thrown with the message:

[className] did not have a zero-argument constructor or a single-argument constructor that accepts SparkConf. Note: if the class is defined inside of another Scala class, then its constructors may accept an implicit parameter that references the enclosing class; in this case, you must define the listener as a top-level class in order to prevent this extra parameter from breaking Spark's ability to find a valid constructor.

Any exception while registering a SparkListenerInterface listener stops the SparkContext and a SparkException is thrown and the source exception’s message.

Exception when registering SparkListener

Set INFO on org.apache.spark.SparkContext logger to see the extra listeners being registered.

INFO SparkContext: Registered listener pl.japila.spark.CustomSparkListener

Creating SparkEnv for Driver — createSparkEnv Method

  conf: SparkConf,
  isLocal: Boolean,
  listenerBus: LiveListenerBus): SparkEnv

createSparkEnv simply delegates the call to SparkEnv to create a SparkEnv for the driver.

It calculates the number of cores to 1 for local master URL, the number of processors available for JVM for * or the exact number in the master URL, or 0 for the cluster master URLs.

Utils.getCurrentUserName Method

getCurrentUserName(): String

getCurrentUserName computes the user name who has started the SparkContext instance.

It is later available as SparkContext.sparkUser.

Internally, it reads SPARK_USER environment variable and, if not set, reverts to Hadoop Security API’s UserGroupInformation.getCurrentUser().getShortUserName().

It is another place where Spark relies on Hadoop API for its operation.

Utils.localHostName Method

localHostName computes the local host name.

It starts by checking SPARK_LOCAL_HOSTNAME environment variable for the value. If it is not defined, it uses SPARK_LOCAL_IP to find the name (using InetAddress.getByName). If it is not defined either, it calls InetAddress.getLocalHost for the name.

Utils.localHostName is executed while SparkContext is created and also to compute the default value of Spark property.
FIXME Review the rest.

stopped Flag

FIXME Where is this used?