Master is the manager of a Spark Standalone cluster.
Master can be launched from command line.
Master RPC Endpoint¶
Launching Standalone Master¶
Master can be launched as a standalone application using
main Entry Point¶
main( argStrings: Array[String]): Unit
main is the entry point of the
Master standalone application.
main prints out the following INFO message to the logs:
Started daemon with process name: [processName]
main registers signal handlers for
In the end,
main requests the
RpcEnv to be notified when
Master supports command-line options.
Usage: Master [options] Options: -i HOST, --ip HOST Hostname to listen on (deprecated, please use --host or -h) -h HOST, --host HOST Hostname to listen on -p PORT, --port PORT Port to listen on (default: 7077) --webui-port PORT Port for web UI (default: 8080) --properties-file FILE Path to a custom Spark properties file. Default is conf/spark-defaults.conf.
Master takes the following to be created:
- web UI's Port
Master is created when:
Masterutility is requested to start up RPC environment
Starting Up RPC Environment¶
startRpcEnvAndEndpoint( host: String, port: Int, webUiPort: Int, conf: SparkConf): (RpcEnv, Int, Option[Int])
startRpcEnvAndEndpoint creates a
RpcEnv with sparkMaster name (and the input arguments) and registers Master endpoint with Master name.
startRpcEnvAndEndpoint is used when:
Scheduling Resources Among Waiting Applications¶
schedule is used when:
Masteris requested to schedule resources among waiting applications
MasterWebUI is the Web UI server for the standalone master. Master starts Web UI to listen to
http://[master's hostname]:webUIPort (e.g.
Successfully started service 'MasterUI' on port 8080. Started MasterWebUI at http://192.168.1.4:8080
Master can be in the following states:
STANDBY- the initial state while
ALIVE- start scheduling resources among applications
To be Reviewed¶
Application ids follows the pattern
Master can be <
The standalone Master starts the REST Server service for alternative application submission that is supposed to work across Spark versions. It is enabled by default (see <
RestSubmissionClient is the client.
The server includes a JSON representation of
SubmitRestProtocolResponse in the HTTP body.
The following INFOs show up when the Master Endpoint starts up (
Master#onStart is called) with REST Server enabled:
INFO Utils: Successfully started service on port 6066. INFO StandaloneRestServer: Started REST server for submitting applications on port 6066
A standalone Master can run with recovery mode enabled and be able to recover state among the available swarm of masters. By default, there is no recovery, i.e. no persistence and no election.
NOTE: Only a master can schedule tasks so having one always on is important for cases where you want to launch new tasks. Running tasks are unaffected by the state of the master.
spark.deploy.recoveryMode to set up the recovery mode (see <
The Recovery Mode enables <
TIP: Check out the exercise link:exercises/spark-exercise-standalone-master-ha.md[Spark Standalone - Using ZooKeeper for High-Availability of Master].
Master communicates with drivers, executors and configures itself using RPC messages.
The following message types are accepted by master (see
A RegisterApplication event is sent by link:spark-standalone.md#AppClient[AppClient] to the standalone Master. The event holds information about the application being deployed (
ApplicationDescription) and the driver's endpoint reference.
ApplicationDescription describes an application by its name, maximum number of cores, executor's memory, command, appUiUrl, and user with optional eventLogDir and eventLogCodec for Event Logs, and the number of cores per executor.
CAUTION: FIXME Finish
A standalone Master receives
RegisterApplication with a
ApplicationDescription and the driver's xref:rpc:RpcEndpointRef.md[RpcEndpointRef].
INFO Registering app " + description.name
Application ids in Spark Standalone are in the format of
Master keeps track of the number of already-scheduled applications (
ApplicationDescription (AppClient) → ApplicationInfo (Master) - application structure enrichment
ApplicationSource metrics +
INFO Registered app " + description.name + " with ID " + app.id
schedule() schedules the currently available resources among waiting apps.
FIXME When is
schedule() method called?
It's only executed when the Master is in
WorkerState.ALIVE state can accept applications.
A driver has a state, i.e.
driver.state and when it's in
DriverState.RUNNING state the driver has been assigned to a worker for execution.
LaunchDriver RPC message¶
WARNING: It seems a dead message. Disregard it for now.
A LaunchDriver message is sent by an active standalone Master to a worker to launch a driver.
.Master finds a place for a driver (posts LaunchDriver) image::spark-standalone-master-worker-LaunchDriver.png[align="center"]
You should see the following INFO in the logs right before the message is sent out to a worker:
INFO Launching driver [driver.id] on worker [worker.id]
The message holds information about the id and name of the driver.
A driver can be running on a single worker while a worker can have many drivers running.
When a worker receives a
LaunchDriver message, it prints out the following INFO:
Asked to launch driver [driver.id]
It then creates a
DriverRunner and starts it. It starts a separate JVM process.
Workers' free memory and cores are considered when assigning some to waiting drivers (applications).
CAUTION: FIXME Go over
Internals of org.apache.spark.deploy.master.Master¶
Master starts, it first creates the default
SparkConf configuration whose values it then overrides using <
A fully-configured master instance requires
8080) settings defined.
TIP: When in troubles, consult link:spark-tips-and-tricks.md[Spark Tips and Tricks] document.
It starts <
master-forward-message-thread to schedule a thread every
spark.worker.timeout to check workers' availability and remove timed-out workers.
It is that Master sends
CheckForWorkerTimeOut message to itself to trigger verification.
When a worker hasn't responded for
spark.worker.timeout, it is assumed dead and the following WARN message appears in the logs:
WARN Removing [worker.id] because we got no heartbeat in [spark.worker.timeout] seconds
System Environment Variables¶
Master uses the following system environment variables (directly or indirectly):
SPARK_LOCAL_HOSTNAME- the custom host name
SPARK_LOCAL_IP- the custom IP to use when
SPARK_LOCAL_HOSTNAMEis not set
SPARK_MASTER_IPas used in
start-master.shscript above!) - the master custom host
7077) - the master custom port
8080) - the port of the master's WebUI. Overriden by
spark.master.ui.portif set in the properties file.
SPARK_PUBLIC_DNS(default: hostname) - the custom master hostname for WebUI's http URL and master's address.
$SPARK_HOME/conf) - the directory of the default properties file link:spark-properties.md#spark-defaults-conf[spark-defaults.conf] from which all properties that start with
spark.prefix are loaded.
Master uses the following properties:
0) - total expected number of cores. When set, an application could get executors of different sizes (in terms of cores).
NONE) - possible modes:
CUSTOM. Refer to <
spark.deploy.recoveryMode.factory- the class name of the custom
spark.deploy.recoveryDirectory(default: empty) - the directory to persist recovery state
- link:spark-standalone.md#spark.deploy.spreadOut[spark.deploy.spreadOut] to perform link:spark-standalone.md#round-robin-scheduling[round-robin scheduling across the nodes].
Int.MaxValue, i.e. unbounded) - the number of maxCores for applications that don't specify it.
60) - time (in seconds) when no heartbeat from a worker means it is lost. See <