Spark Standalone Mode

2020-01-21
  • This will become a table of contents (this text will be scraped). {:toc}

In addition to running on the Mesos or YARN cluster managers, Spark also provides a simple standalone deploy mode. You can launch a standalone cluster either manually, by starting a master and workers by hand, or use our provided launch scripts. It is also possible to run these daemons on a single machine for testing.

Security

Security in Spark is OFF by default. This could mean you are vulnerable to attack by default. Please see Spark Security and the specific security sections in this doc before running Spark.

Installing Spark Standalone to a Cluster

To install Spark Standalone mode, you simply place a compiled version of Spark on each node on the cluster. You can obtain pre-built versions of Spark with each release or build it yourself.

Starting a Cluster Manually

You can start a standalone master server by executing:

./sbin/start-master.sh

Once started, the master will print out a spark://HOST:PORT URL for itself, which you can use to connect workers to it, or pass as the "master" argument to SparkContext. You can also find this URL on the master's web UI, which is http://localhost:8080 by default.

Similarly, you can start one or more workers and connect them to the master via:

./sbin/start-slave.sh <master-spark-URL>

Once you have started a worker, look at the master's web UI (http://localhost:8080 by default). You should see the new node listed there, along with its number of CPUs and memory (minus one gigabyte left for the OS).

Finally, the following configuration options can be passed to the master and worker:

Argument Meaning
-h HOST, --host HOST Hostname to listen on
-i HOST, --ip HOST Hostname to listen on (deprecated, use -h or --host)
-p PORT, --port PORT Port for service to listen on (default: 7077 for master, random for worker)
--webui-port PORT Port for web UI (default: 8080 for master, 8081 for worker)
-c CORES, --cores CORES Total CPU cores to allow Spark applications to use on the machine (default: all available); only on worker
-m MEM, --memory MEM Total amount of memory to allow Spark applications to use on the machine, in a format like 1000M or 2G (default: your machine's total RAM minus 1 GiB); only on worker
-d DIR, --work-dir DIR Directory to use for scratch space and job output logs (default: SPARK_HOME/work); only on worker
--properties-file FILE Path to a custom Spark properties file to load (default: conf/spark-defaults.conf)

Cluster Launch Scripts

To launch a Spark standalone cluster with the launch scripts, you should create a file called conf/slaves in your Spark directory, which must contain the hostnames of all the machines where you intend to start Spark workers, one per line. If conf/slaves does not exist, the launch scripts defaults to a single machine (localhost), which is useful for testing. Note, the master machine accesses each of the worker machines via ssh. By default, ssh is run in parallel and requires password-less (using a private key) access to be setup. If you do not have a password-less setup, you can set the environment variable SPARK_SSH_FOREGROUND and serially provide a password for each worker.

Once you've set up this file, you can launch or stop your cluster with the following shell scripts, based on Hadoop's deploy scripts, and available in SPARK_HOME/sbin:

  • sbin/start-master.sh - Starts a master instance on the machine the script is executed on.
  • sbin/start-slaves.sh - Starts a worker instance on each machine specified in the conf/slaves file.
  • sbin/start-slave.sh - Starts a worker instance on the machine the script is executed on.
  • sbin/start-all.sh - Starts both a master and a number of workers as described above.
  • sbin/stop-master.sh - Stops the master that was started via the sbin/start-master.sh script.
  • sbin/stop-slave.sh - Stops all worker instances on the machine the script is executed on.
  • sbin/stop-slaves.sh - Stops all worker instances on the machines specified in the conf/slaves file.
  • sbin/stop-all.sh - Stops both the master and the workers as described above.

Note that these scripts must be executed on the machine you want to run the Spark master on, not your local machine.

You can optionally configure the cluster further by setting environment variables in conf/spark-env.sh. Create this file by starting with the conf/spark-env.sh.template, and copy it to all your worker machines for the settings to take effect. The following settings are available:

Environment Variable Meaning
SPARK_MASTER_HOST Bind the master to a specific hostname or IP address, for example a public one.
SPARK_MASTER_PORT Start the master on a different port (default: 7077).
SPARK_MASTER_WEBUI_PORT Port for the master web UI (default: 8080).
SPARK_MASTER_OPTS Configuration properties that apply only to the master in the form "-Dx=y" (default: none). See below for a list of possible options.
SPARK_LOCAL_DIRS Directory to use for "scratch" space in Spark, including map output files and RDDs that get stored on disk. This should be on a fast, local disk in your system. It can also be a comma-separated list of multiple directories on different disks.
SPARK_WORKER_CORES Total number of cores to allow Spark applications to use on the machine (default: all available cores).
SPARK_WORKER_MEMORY Total amount of memory to allow Spark applications to use on the machine, e.g. 1000m, 2g (default: total memory minus 1 GiB); note that each application's individual memory is configured using its spark.executor.memory property.
SPARK_WORKER_PORT Start the Spark worker on a specific port (default: random).
SPARK_WORKER_WEBUI_PORT Port for the worker web UI (default: 8081).
SPARK_WORKER_DIR Directory to run applications in, which will include both logs and scratch space (default: SPARK_HOME/work).
SPARK_WORKER_OPTS Configuration properties that apply only to the worker in the form "-Dx=y" (default: none). See below for a list of possible options.
SPARK_DAEMON_MEMORY Memory to allocate to the Spark master and worker daemons themselves (default: 1g).
SPARK_DAEMON_JAVA_OPTS JVM options for the Spark master and worker daemons themselves in the form "-Dx=y" (default: none).
SPARK_DAEMON_CLASSPATH Classpath for the Spark master and worker daemons themselves (default: none).
SPARK_PUBLIC_DNS The public DNS name of the Spark master and workers (default: none).

Note: The launch scripts do not currently support Windows. To run a Spark cluster on Windows, start the master and workers by hand.

SPARK_MASTER_OPTS supports the following system properties:

Property Name Default Meaning
spark.deploy.retainedApplications 200 The maximum number of completed applications to display. Older applications will be dropped from the UI to maintain this limit.
spark.deploy.retainedDrivers 200 The maximum number of completed drivers to display. Older drivers will be dropped from the UI to maintain this limit.
spark.deploy.spreadOut true Whether the standalone cluster manager should spread applications out across nodes or try to consolidate them onto as few nodes as possible. Spreading out is usually better for data locality in HDFS, but consolidating is more efficient for compute-intensive workloads.
spark.deploy.defaultCores (infinite) Default number of cores to give to applications in Spark's standalone mode if they don't set spark.cores.max. If not set, applications always get all available cores unless they configure spark.cores.max themselves. Set this lower on a shared cluster to prevent users from grabbing the whole cluster by default.
spark.deploy.maxExecutorRetries 10 Limit on the maximum number of back-to-back executor failures that can occur before the standalone cluster manager removes a faulty application. An application will never be removed if it has any running executors. If an application experiences more than spark.deploy.maxExecutorRetries failures in a row, no executors successfully start running in between those failures, and the application has no running executors then the standalone cluster manager will remove the application and mark it as failed. To disable this automatic removal, set spark.deploy.maxExecutorRetries to -1.
spark.worker.timeout 60 Number of seconds after which the standalone deploy master considers a worker lost if it receives no heartbeats.
spark.worker.resource.{resourceName}.amount (none) Amount of a particular resource to use on the worker.
spark.worker.resource.{resourceName}.discoveryScript (none) Path to resource discovery script, which is used to find a particular resource while worker starting up. And the output of the script should be formatted like the ResourceInformation class. When spark.resources.coordinate.enable is off, the discovery script must assign different resources for workers and drivers in client mode that run on the same host to avoid resource conflict.
spark.worker.resourcesFile (none) Path to resources file which is used to find various resources while worker starting up. The content of resources file should be formatted like [[{"id":{"componentName": "spark.worker","resourceName":"gpu"},"addresses":["0","1","2"]}]]. When spark.resources.coordinate.enable is off, resources file must assign different resources for workers and drivers in client mode that run on the same host to avoid resource conflict. If a particular resource is not found in the resources file, the discovery script would be used to find that resource. If the discovery script also does not find the resources, the worker will fail to start up.

SPARK_WORKER_OPTS supports the following system properties:

Property Name Default Meaning
spark.worker.cleanup.enabled false Enable periodic cleanup of worker / application directories. Note that this only affects standalone mode, as YARN works differently. Only the directories of stopped applications are cleaned up. This should be enabled if spark.shuffle.service.db.enabled is "true"
spark.worker.cleanup.interval 1800 (30 minutes) Controls the interval, in seconds, at which the worker cleans up old application work dirs on the local machine.
spark.worker.cleanup.appDataTtl 604800 (7 days, 7 * 24 * 3600) The number of seconds to retain application work directories on each worker. This is a Time To Live and should depend on the amount of available disk space you have. Application logs and jars are downloaded to each application work dir. Over time, the work dirs can quickly fill up disk space, especially if you run jobs very frequently.
spark.shuffle.service.db.enabled true Store External Shuffle service state on local disk so that when the external shuffle service is restarted, it will automatically reload info on current executors. This only affects standalone mode (yarn always has this behavior enabled). You should also enable spark.worker.cleanup.enabled, to ensure that the state eventually gets cleaned up. This config may be removed in the future.
spark.storage.cleanupFilesAfterExecutorExit true Enable cleanup non-shuffle files(such as temp. shuffle blocks, cached RDD/broadcast blocks, spill files, etc) of worker directories following executor exits. Note that this doesn't overlap with `spark.worker.cleanup.enabled`, as this enables cleanup of non-shuffle files in local directories of a dead executor, while `spark.worker.cleanup.enabled` enables cleanup of all files/subdirectories of a stopped and timeout application. This only affects Standalone mode, support of other cluster manangers can be added in the future.
spark.worker.ui.compressedLogFileLengthCacheSize 100 For compressed log files, the uncompressed file can only be computed by uncompressing the files. Spark caches the uncompressed file size of compressed log files. This property controls the cache size.

Resource Allocation and Configuration Overview

Please make sure to have read the Custom Resource Scheduling and Configuration Overview section on the configuration page. This section only talks about the Spark Standalone specific aspects of resource scheduling.

Spark Standalone has 2 parts, the first is configuring the resources for the Worker, the second is the resource allocation for a specific application.

The user must configure the Workers to have a set of resources available so that it can assign them out to Executors. The spark.worker.resource.{resourceName}.amount is used to control the amount of each resource the worker has allocated. The user must also specify either spark.worker.resourcesFile or spark.worker.resource.{resourceName}.discoveryScript to specify how the Worker discovers the resources its assigned. See the descriptions above for each of those to see which method works best for your setup. Please take note of spark.resources.coordinate.enable as it indicates whether Spark should handle coordinating resources or if the user has made sure each Worker has separate resources. Also note that if using the resources coordination spark.resources.dir can be used to specify the directory used to do that coordination.

The second part is running an application on Spark Standalone. The only special case from the standard Spark resource configs is when you are running the Driver in client mode. For a Driver in client mode, the user can specify the resources it uses via spark.driver.resourcesfile or spark.driver.resources.{resourceName}.discoveryScript. If the Driver is running on the same host as other Drivers or Workers there are 2 ways to make sure the they don't use the same resources. The user can either configure spark.resources.coordinate.enable on and give all the Driver/Workers the same set or resources and Spark will handle make sure each Driver/Worker has separate resources, or the user can make sure the resources file or discovery script only returns resources the do not conflict with other Drivers or Workers running on the same node.

Note, the user does not need to specify a discovery script when submitting an application as the Worker will start each Executor with the resources it allocates to it.

Connecting an Application to the Cluster

To run an application on the Spark cluster, simply pass the spark://IP:PORT URL of the master as to the SparkContext constructor.

To run an interactive Spark shell against the cluster, run the following command:

./bin/spark-shell --master spark://IP:PORT

You can also pass an option --total-executor-cores <numCores> to control the number of cores that spark-shell uses on the cluster.

Launching Spark Applications

The spark-submit script provides the most straightforward way to submit a compiled Spark application to the cluster. For standalone clusters, Spark currently supports two deploy modes. In client mode, the driver is launched in the same process as the client that submits the application. In cluster mode, however, the driver is launched from one of the Worker processes inside the cluster, and the client process exits as soon as it fulfills its responsibility of submitting the application without waiting for the application to finish.

If your application is launched through Spark submit, then the application jar is automatically distributed to all worker nodes. For any additional jars that your application depends on, you should specify them through the --jars flag using comma as a delimiter (e.g. --jars jar1,jar2). To control the application's configuration or execution environment, see Spark Configuration.

Additionally, standalone cluster mode supports restarting your application automatically if it exited with non-zero exit code. To use this feature, you may pass in the --supervise flag to spark-submit when launching your application. Then, if you wish to kill an application that is failing repeatedly, you may do so through:

./bin/spark-class org.apache.spark.deploy.Client kill <master url> <driver ID>

You can find the driver ID through the standalone Master web UI at http://<master url>:8080.

Resource Scheduling

The standalone cluster mode currently only supports a simple FIFO scheduler across applications. However, to allow multiple concurrent users, you can control the maximum number of resources each application will use. By default, it will acquire all cores in the cluster, which only makes sense if you just run one application at a time. You can cap the number of cores by setting spark.cores.max in your SparkConf. For example:

val conf = new SparkConf() .setMaster(...) .setAppName(...) .set("spark.cores.max", "10") val sc = new SparkContext(conf)

In addition, you can configure spark.deploy.defaultCores on the cluster master process to change the default for applications that don't set spark.cores.max to something less than infinite. Do this by adding the following to conf/spark-env.sh:

export SPARK_MASTER_OPTS="-Dspark.deploy.defaultCores= "

This is useful on shared clusters where users might not have configured a maximum number of cores individually.

Executors Scheduling

The number of cores assigned to each executor is configurable. When spark.executor.cores is explicitly set, multiple executors from the same application may be launched on the same worker if the worker has enough cores and memory. Otherwise, each executor grabs all the cores available on the worker by default, in which case only one executor per application may be launched on each worker during one single schedule iteration.

Monitoring and Logging

Spark's standalone mode offers a web-based user interface to monitor the cluster. The master and each worker has its own web UI that shows cluster and job statistics. By default, you can access the web UI for the master at port 8080. The port can be changed either in the configuration file or via command-line options.

In addition, detailed log output for each job is also written to the work directory of each slave node (SPARK_HOME/work by default). You will see two files for each job, stdout and stderr, with all output it wrote to its console.

Running Alongside Hadoop

You can run Spark alongside your existing Hadoop cluster by just launching it as a separate service on the same machines. To access Hadoop data from Spark, just use an hdfs:// URL (typically hdfs://<namenode>:9000/path, but you can find the right URL on your Hadoop Namenode's web UI). Alternatively, you can set up a separate cluster for Spark, and still have it access HDFS over the network; this will be slower than disk-local access, but may not be a concern if you are still running in the same local area network (e.g. you place a few Spark machines on each rack that you have Hadoop on).

Configuring Ports for Network Security

Generally speaking, a Spark cluster and its services are not deployed on the public internet. They are generally private services, and should only be accessible within the network of the organization that deploys Spark. Access to the hosts and ports used by Spark services should be limited to origin hosts that need to access the services.

This is particularly important for clusters using the standalone resource manager, as they do not support fine-grained access control in a way that other resource managers do.

For a complete list of ports to configure, see the security page.

High Availability

By default, standalone scheduling clusters are resilient to Worker failures (insofar as Spark itself is resilient to losing work by moving it to other workers). However, the scheduler uses a Master to make scheduling decisions, and this (by default) creates a single point of failure: if the Master crashes, no new applications can be created. In order to circumvent this, we have two high availability schemes, detailed below.

Standby Masters with ZooKeeper

Overview

Utilizing ZooKeeper to provide leader election and some state storage, you can launch multiple Masters in your cluster connected to the same ZooKeeper instance. One will be elected "leader" and the others will remain in standby mode. If the current leader dies, another Master will be elected, recover the old Master's state, and then resume scheduling. The entire recovery process (from the time the first leader goes down) should take between 1 and 2 minutes. Note that this delay only affects scheduling new applications -- applications that were already running during Master failover are unaffected.

Learn more about getting started with ZooKeeper here.

Configuration

In order to enable this recovery mode, you can set SPARK_DAEMON_JAVA_OPTS in spark-env by configuring spark.deploy.recoveryMode and related spark.deploy.zookeeper.* configurations. For more information about these configurations please refer to the configuration doc

Possible gotcha: If you have multiple Masters in your cluster but fail to correctly configure the Masters to use ZooKeeper, the Masters will fail to discover each other and think they're all leaders. This will not lead to a healthy cluster state (as all Masters will schedule independently).

Details

After you have a ZooKeeper cluster set up, enabling high availability is straightforward. Simply start multiple Master processes on different nodes with the same ZooKeeper configuration (ZooKeeper URL and directory). Masters can be added and removed at any time.

In order to schedule new applications or add Workers to the cluster, they need to know the IP address of the current leader. This can be accomplished by simply passing in a list of Masters where you used to pass in a single one. For example, you might start your SparkContext pointing to spark://host1:port1,host2:port2. This would cause your SparkContext to try registering with both Masters -- if host1 goes down, this configuration would still be correct as we'd find the new leader, host2.

There's an important distinction to be made between "registering with a Master" and normal operation. When starting up, an application or Worker needs to be able to find and register with the current lead Master. Once it successfully registers, though, it is "in the system" (i.e., stored in ZooKeeper). If failover occurs, the new leader will contact all previously registered applications and Workers to inform them of the change in leadership, so they need not even have known of the existence of the new Master at startup.

Due to this property, new Masters can be created at any time, and the only thing you need to worry about is that new applications and Workers can find it to register with in case it becomes the leader. Once registered, you're taken care of.

Single-Node Recovery with Local File System

Overview

ZooKeeper is the best way to go for production-level high availability, but if you just want to be able to restart the Master if it goes down, FILESYSTEM mode can take care of it. When applications and Workers register, they have enough state written to the provided directory so that they can be recovered upon a restart of the Master process.

Configuration

In order to enable this recovery mode, you can set SPARK_DAEMON_JAVA_OPTS in spark-env using this configuration:

System property Meaning
spark.deploy.recoveryMode Set to FILESYSTEM to enable single-node recovery mode (default: NONE).
spark.deploy.recoveryDirectory The directory in which Spark will store recovery state, accessible from the Master's perspective.

Details

  • This solution can be used in tandem with a process monitor/manager like monit, or just to enable manual recovery via restart.
  • While filesystem recovery seems straightforwardly better than not doing any recovery at all, this mode may be suboptimal for certain development or experimental purposes. In particular, killing a master via stop-master.sh does not clean up its recovery state, so whenever you start a new Master, it will enter recovery mode. This could increase the startup time by up to 1 minute if it needs to wait for all previously-registered Workers/clients to timeout.
  • While it's not officially supported, you could mount an NFS directory as the recovery directory. If the original Master node dies completely, you could then start a Master on a different node, which would correctly recover all previously registered Workers/applications (equivalent to ZooKeeper recovery). Future applications will have to be able to find the new Master, however, in order to register.